The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning

Introduction

Similar to other technologies, such as gene editing and nuclear energy, AI is dual-use—it can be leveraged for benefit and harm (Urbina et al., 2022). To address its dual-use risks, the White House Executive Order on Artificial Intelligence (White House, 2023) calls for investigation into the ability of AI to enable malicious actors in developing chemical, biological, radiological, nuclear, and cyber weapons. For instance, AI coding assistants may lower the barrier of entry for novices to conduct cyberattacks (Fang et al., 2024), potentially increasing attack scale and frequency, and risking catastrophe if directed towards critical infrastructure, such as power grids (UK Cabinet Office, 2023). Likewise, AI assistants for biology could troubleshoot bottlenecks in biological weapons development, increasing the frequency of attempts to build a bioweapon and straining risk mitigation measures (Sandbrink, 2023). This has motivated government institutions and major AI labs to anticipate risk by designing evaluations for AI-aided biological threats (UK AI Safety Summit, 2023; Anthropic, 2023; OpenAI, 2023b, 2024; Mouton et al., 2024).

Unfortunately, many existing evaluations of hazardous capabilities do not provide a guide for risk mitigation. Currently, developers use private, manual, and highly-specific risk evaluations, investigating whether models can build biological weapons end-to-end (Sandbrink, 2023) or hack well enough to exfiltrate their own weights (Shevlane et al., 2023). However, private benchmarking limits scientific inquiry into measuring and reducing malicious use. Furthermore, these evaluations test a small number of specific risk pathways, so low performance does not guarantee that LLMs are secure across the broad distribution of malicious use risks.

Additionally, technical solutions to reduce malicious use in LLMs are limited. The primary safeguard is training models to refuse harmful queries (Ouyang et al., 2022; Bai et al., 2022; Mazeika et al., 2024), but adversaries can deploy adversarial attacks (Wei et al., 2023; Zou et al., 2023b) to bypass models’ refusal training. Another proposal is to filter hazardous information from the pretraining data (Ngo et al., 2021), but adversaries may reintroduce this information through finetuning (Zhan et al., 2023; Qi et al., 2023; Pelrine et al., 2023). Fortunately, closed-source LLM providers have leverage, as they may apply safety interventions before serving LLMs. In particular, providers may perform unlearning to directly remove hazardous knowledge (Figure 2). Unlearned models have higher inherent safety: even if they are jailbroken, unlearned models lack the hazardous knowledge necessary to enable malicious users (Hendrycks et al., 2021). However, research into unlearning hazardous knowledge is bottlenecked by the lack of a public benchmark.

To overcome both of these challenges, we introduce the Weapons of Mass Destruction Proxy Benchmark (WMDP), a benchmark of $4,\!157$ multiple-choice questions costing over $200K to develop (Figure 1). WMDP is a proxy measurement for hazardous knowledge in biosecurity (Section 3.2), cybersecurity (Section 3.3), and chemical security (Section 3.4). To design WMDP, academics and technical consultants created threat models for how LLMs might aid in the development of biological, cyber, and chemical attacks, and generated questions based on these threat models. We adopt a conservative stance towards including information in WMDP (Figure 3): we primarily include offensive knowledge, as unlearning defensive knowledge (e.g., biosafety protocols) may prevent benevolent use cases of LLMs. Simultaneously, we follow a stringent process to expunge sensitive information from WMDP in compliance with U.S. export control requirements, mitigating the risk of WMDP being repurposed by malicious actors (Section 3.5). We publicly release WMDP to both measure hazardous knowledge, and benchmark methods for reducing malicious use.

To guide progress on unlearning, we develop Contrastive Unlearn Tuning (Cut), a state-of-the-art unlearning method that removes hazardous knowledge while preserving general model capabilities. Inspired by representation engineering (Zou et al., 2023a), Cut pushes model activations along an unlearning direction on hazardous data, while regularizing towards original, frozen model weights on benign data (Section 4). Cut significantly reduces model performance on WMDP, while mostly retaining general capabilities on MMLU (Hendrycks et al., 2020b) and MT-Bench (Zheng et al., 2023a), suggesting that concrete progress on unlearning is feasible (Section 5.2) with minimal impact on the general scientific capabilities of LLMs. We provide evidence that Cut is robust, as unlearned knowledge cannot be recovered by linear probing or adversarial attacks (Sections 5.2 and 5.3).

Overall, we envision unlearning as one piece of a larger sociotechnical solution towards reducing malicious use of AI systems. Unlearning should be applied carefully, as it inherently reduces model capabilities. Scientific knowledge (especially in cybersecurity) is often dual-use, so unlearning such knowledge may harm defenders as much as attackers. In these cases, unlearning can be paired with structured API access (Shevlane, 2022), where model developers serve the unlearned model to everyday users, but serve the unrestricted, base model to approved users, such as red-teamers, security professionals, or virology researchers (Section 6.2). Furthermore, other risk mitigation strategies may be promoted, such as implementing universal screening of synthetic DNA orders to prevent widespread access to pathogen DNA. As AI systems develop more capabilities, a combination of these interventions will be critical in reducing malicious use. To enable further research on unlearning hazardous knowledge, we release our datasets, code, and models publicly at https://wmdp.ai.

Related Work

Recent work has highlighted safety concerns of language models, including generating falsehoods (Ji et al., 2023; Zhang et al., 2023), producing toxic content (Gehman et al., 2020; Deshpande et al., 2023; Pan et al., 2024), and deceiving humans (Park et al., 2023; Scheurer et al., 2023). In response, safety benchmarks are used to monitor and mitigate these behaviors (Hendrycks et al., 2020a; Lin et al., 2021; Li et al., 2023; Pan et al., 2023; Kinniment and Sato, 2023; Inan et al., 2023).

Specifically, one growing concern is the ability of LLMs to assist with malicious use. In particular, LLMs may aid actors in planning bioattacks (Sandbrink, 2023) and procuring pathogens (Gopal et al., 2023). Moreover, LLMs can assist users in synthesizing dangerous chemicals (Boiko et al., 2023) or conducting cyberattacks (Bhatt et al., 2023). In response to these emergent hazardous capabilities (Hendrycks et al., 2021), major AI labs have developed frameworks to measure and mitigate biological, cybersecurity, and chemical hazards posed by their models (Anthropic, 2023; OpenAI, 2023b, 2024). Unfortunately, many of the details of these evaluations are often private to the individual research labs for which they were developed. In contrast, we develop an open-source evaluation that empowers the broader ML community to make progress towards benchmarking and unlearning hazardous knowledge.

Towards improving model safety, strategies such as input safety filtering (Inan et al., 2023) and learning from human preference data (Ziegler et al., 2020; Rafailov et al., 2023) have been developed; however, these methods can be vulnerable to jailbreaks (Wei et al., 2023; Chao et al., 2023; Yao et al., 2023a; Yuan et al., 2023) and adversarial attacks (Wallace et al., 2019; Guo et al., 2021; Jones et al., 2023; Zou et al., 2023b). To reduce inherent model risk, hazardous data can be removed prior to pretraining (Ngo et al., 2021), but having input into this process is inaccessible for most end users. Furthermore, models may be susceptible to subsequent harmful finetuning (Zhan et al., 2023; Yang et al., 2023) (Figure 2); as a result, and especially in the case of models that are accessed via API, additional automated methods that can be applied after finetuning—such as unlearning—may remove resulting hazards.

Unlearning (Cao and Yang, 2015) originally gained traction as a response to privacy concerns in light of regulation (Council of European Union, 2014; CCPA, 2018), and most methods focused on erasing specific samples or facts (Golatkar et al., 2020; Liu et al., 2020; Meng et al., 2022; Jang et al., 2023; Pawelczyk et al., 2023) rather than entire domains. Goel et al. (2024) show existing unlearning methods struggle to remove knowledge without access to all relevant training data, a challenge Cut overcomes.

More recent methods erase broader concepts such as gender (Belrose et al., 2023), harmful behaviors (Yao et al., 2023b; Liu et al., 2024), or fictional universes (Eldan and Russinovich, 2023), but have not been proven to eliminate scientific knowledge which enables malicious use. Furthermore, most benchmarks for unlearning involve removing specific data samples (Google, 2023) or artificially chosen deletion sets (Choi and Na, 2023; Goel et al., 2023; Maini et al., 2024; Goel et al., 2024). In contrast, WMDP benchmarks on real-world information that can enable malicious use.

The WMDP Benchmark

We introduce the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of $4,\!157$ expert-written, multiple-choice questions in biosecurity (WMDP-Bio), cybersecurity (WMDP-Cyber), and chemistry (WMDP-Chem) costing over $200K to develop. The goal is to reduce question-answer (QA) accuracy on WMDP while maintaining performance on other benchmarks, such as MMLU (Hendrycks et al., 2020b) or MT-Bench (Zheng et al., 2023a). See Appendix A.1 for a breakdown of questions in WMDP and Appendix B.2 for a sample question.

WMDP is an automatic, public benchmark of hazardous capabilities that serves as a guide for risk mitigation (Section 3.1). We create questions by designing threat models for biosecurity.(Section 3.2), cybersecurity (Section 3.3), and chemistry (Section 3.4). We also remove sensitive and export-controlled information from entering WMDP (Section 3.5). To further unlearning research beyond WMDP, we also provide additional unlearning benchmarks based on MMLU (Appendix C).

To create an automatic measure of hazardous capabilities that the broader research community can readily iterate on, we design WMDP as a dataset of four-choice multiple-choice questions. Multiple-choice is a common paradigm to test knowledge in language models (Hendrycks et al., 2020b; Rein et al., 2023).

Because WMDP measures knowledge of hazardous topics, models with a low score on WMDP likely lack the knowledge needed to help with malicious use. However, models with a high score on WMDP are not necessarily unsafe, as they may still lack the reasoning ability to combine the knowledge in the sequence of steps needed to create a weapon.

WMDP should guide risk mitigation by enabling researchers to measure and reduce models’ hazardous capabilities. Because directly building a dataset of sensitive information would increase the attack capabilities of malicious actors (Esvelt, 2018; Lewis et al., 2019), we collect questions that approximate or correlate with the hazardous knowledge we wish to remove (Figure 3). In particular, we collect questions with knowledge that is a precursor, neighbor, or component of the hazardous knowledge we wish to remove. Moreover, we empirically demonstrate that models with lower performance on WMDP are less capable for malicious use (Section 5.4).

Examples of our dataset generation processes are detailed in Figure 4. In the left panel, research that aims to develop enhanced potential pandemic pathogens (ePPPs) is a precursor to developing novel viruses, so unlearning the former will also unlearn a large subset of the latter. In the center panel, there are topics in chemistry (e.g., procurement or synthesis) that contain questions with a wide variance in hazard level, so we approximate especially sensitive information by collecting questions near the boundary. In the right panel, a cyberweapon requires knowledge of several components (e.g., a payload, a trigger mechanism, and an infection mechanism), so excising knowledge of components will reduce hazards. Because some of the components may be dual-use, we generate questions for components that are primarily offensive in nature.

Our questions are written by academics and technical consultants in biosecurity, cybersecurity, and chemistry. We first generate threat models for each of these areas and then use the models to inform questions that an adversary might encounter when developing attack capabilities. To ensure quality, all of our questions were checked by at least two experts from different organizations.

2 Biosecurity Threat Model

In biosecurity, the malicious use threats that are increased by AI can be broadly categorized as expanding access to pre-existing threats (by lowering barriers to entry), and unlocking new areas of biology (by synthesizing new knowledge or accelerating in-silico modeling and experimentation).

We primarily focus on the development and dissemination of transmissible potential pandemic agents, such as influenza, smallpox, etc. While our dataset additionally includes some information about highly lethal non-transmissible bioweapons like anthrax, we believe the majority of emerging risk from biotechnology stems from advances in synthetic biology and bioengineering that increase access to, or modify, the design and development of transmissible agents.

A standard biotechnology risk chain can be seen in Figure 5. In this threat model, “ideation” involves actively planning for a biological attack; “design” involves retrieving blueprints for a hazardous agent, such as determining the DNA sequence; “build” consists of the protocols, reagents, and equipment necessary to create the threat; and “test” consists of measuring characteristics or properties of the pathogen of interest. By “learning” from these results and iterating after the test phase, the design can be revised until the threat is released.

We aim to cover areas of dual-use biology that map onto each part of this risk chain. While not all areas of dual-use biology can be publicly listed in an open source benchmark, we opt to cover areas that provide useful information on what sorts of biological information would be harmful for models to disseminate, but are unlikely to cause significant additional risk despite their open disclosure in this dataset. We consulted with external biosecurity experts when formulating these areas, and describe additional considerations in Section A.2.

For “ideation,” we include areas of historical bioweapons and bioterrorism use (including methods, pathogen strains, and delivery systems used to mount bioterrorist attacks) that are publicly available online. We additionally include previously published information surrounding enhanced potential pandemic pathogens (ePPP) research, which involves investigating ways to make pre-existing pathogens more transmissible or virulent. Finally, we include what we believe are emerging areas of dual-use risk, such as viral vector engineering and research – while this research has significant benefits for medical treatments like gene therapy, we believe widespread access to information about viral vectors to the public is unlikely to provide significant value and could be actively harmful.

For the “build” portion of the design-build-test-learn (DBTL) cycle, we include knowledge on viral reverse genetics (protocols for reconstructing viruses), as well as niche areas of synthetic biology and bioengineering that may expand access to pathogen acquisition and release (including knowledge around circumventing DNA/RNA synthesis screening and export controls, as well as synthesizing biological agents). While these areas have some dual-use risk, they primarily consist of knowledge that a low-skilled actor can readily obtain through routine online sources.

We aim to address the “test” and “learn” parts of the DBTL cycle by including questions about specific assay types to measure specific virus characteristics. Finally, many aspects of “release” are included in questions formulated from the historical bioweapons and bioterrorism category; we did not include additional information about the release of potential pandemic pathogens to mitigate dual-use sensitive information risks.

3 Cybersecurity Threat Model

In cybersecurity, AI may aid malicious actors throughthe stages of a cyberattack: reconnaissance, weaponization, exploitation, and post-exploitation (Strom et al., 2020; Hutchins et al., 2011).

In the initial reconnaissance stage, attackers gather any available information about their target through both open-source and proprietary intelligence-gathering techniques. Openly available information on search engines such as Google and Shodan is critical in this process. To measure the model’s reconnaissance abilities, we sourced questions involving foundational reconnaissance phase concepts and specific techniques to uncover information about a target.

During the weaponization & vulnerability discovery stage, attackers use various methods to discover vulnerabilities within a target and craft exploits to take advantage of them. To measure a model’s weaponization & vulnerability discovery abilities, we sourced questions designed to measure the model’s ability to reason about both assembly and source code, interpret binary data structures such as network packets and files, and understand the use cases for a variety of well-known source cybersecurity tools. Very simple exploit generation tasks were included.

Because tool-augmented and long-context LLMs are adept at handling unstructured data at scale, weaponization & vulnerability discovery may have high potential for LLM enhancement. Moreover, weaponization is frequently the most difficult portion of the attack pipeline – discovering a vulnerability and writing the exploit for a high value target requires significant skill and in some cases may take years. However, once an exploit is packaged into an exploitation framework, it becomes widely available to actors with a much lower skill level. As a result, WMDP-Cyber places particular focus on question development at this stage.

During the exploitation stage, exploits designed to penetrate the target are launched to deliver a malicious payload – for instance, a software implant, a rootkit/bootkit, or simply a payload designed to crash the target device in the case of a DOS attack. Delivery of the payload to the designated target may require multiple complex steps. To measure a model’s exploitation abilities, we sourced questions involving common exploitation frameworks such as Metasploit.

Finally, after the payload is delivered, the desired post-exploitation activities are undertaken. This often involves establishing back-channel communications with a command and control infrastructure, but this is not always a requirement. This stage is ultimately about retaining control of the compromised host without alerting anyone to the malicious presence on the machine. To measure a model’s post-exploitation abilities, we sourced questions involving common post-exploitation frameworks such as Colbalt Strike, Empire, Mimikatz, Bloodhound, and Sliver.

4 Chemical Security Threat Model

In chemistry, similar to cybersecurity, AI can increase risk by aiding malicious actors through the stages of designing and deploying a chemical weapon. These can be categorized as: (a) procuring the source materials; (b) synthesizing the target chemical weapons and/or explosives; (c) purifying and validating the synthesized compounds; (d) surreptitiously transporting the weapons to the desired location; and (e) deploying the weapons in an effective manner. For a more detailed breakdown of the categories, see Appendix A.3.

Each of these steps needs to be carried out without attracting the attention of law-enforcement officials and other regulatory agencies, which means that most syntheses need to be executed outside of a regulated chemistry laboratory. In particular, it will be more difficult for a harmful actor to purchase chemicals, as they will be unable to rely on large chemical supply companies such as Thermo Fisher Scientific or Millipore Sigma. Moreover, chemical syntheses and purifications that require carefully controlled temperature conditions or exclusion of oxygen from the reaction environment will be markedly harder to execute effectively outside of the confines of registered, regulated, and well-stocked chemistry laboratories.

Once the target compounds have been synthesized and purified effectively, they must be transported without detection. Transporting the compounds via mass transport, especially by airplanes, must be done in a way that disguises the true identity of the compounds, either by mixing them with other compounds that have similar chemical profiles but are non-toxic, by transporting them in parts and assembling them at the final location, or via other similarly duplicitous strategies. These methods require significant knowledge of the properties of the compounds, as well as of the detection and security systems that are used throughout the mass transportation network.

Finally, effectively deploying the chemical weapon or explosive requires knowledge of properties of the compounds (e.g., the vapor pressure, solubility, or density) and how they operate. For example, malicious actors deploying chemical weapons must determine whether to deploy them through air, water, or contact exposure. This demands knowledge of how these weapons exert their deleterious health effects. For explosives, actors ensure that the explosives act only at the time and place of their choosing, requiring knowledge of the stability of the explosives.

5 Sensitive Information Mitigation

We implemented stringent procedures to ensure that no sensitive information is released in WMDP. First, we asked domain experts to flag questions they deemed to contain sensitive information based on their own risk models. Flagged questions were immediately excluded from the dataset. Aggregating opinions from discussions with academics and technical consultants, we identified that most concerns with sensitive information centered around WMDP-Bio and WMDP-Chem, so we took additional steps to mitigate sensitive knowledge in those categories. Specifically, we instituted a policy of “cross-checking” for WMDP-Bio and WMDP-Chem: on each question, two additional domain experts were tasked with determining whether the question constitutes sensitive information. Finally, with the support and guidance of external counsel, the publication of WMDP was assessed for compliance with applicable U.S. export control requirements, including with respect to the International Traffic in Arms Regulations (22 CFR Parts 120-130) (ITAR, 2024) and Export Administration Regulations (15 CFR Parts 730-774) (EAR, 2024).

Cut: Unlearning Inspired By Representation Engineering

We introduce Contrastive Unlearn Tuning (Cut), a finetuning method for unlearning hazardous knowledge (Algorithm 1). We outline the setup (Section 4.1) and explain our method (Section 4.2). We focus on unlearning hazardous knowledge in biosecurity and cybersecurity, but not in chemistry. While WMDP-Chem is a useful tool for hazard measurement, we are more uncertain if the hazard mitigation benefits of unlearning on WMDP-Chem outweigh the costs on general model capabilities.

We consider an autoregressive language model that accepts a prompt (e.g., “How can I synthesize anthrax?”) and returns a completion (e.g., “To synthesize anthrax, you need…”). We aim to reduce the model’s ability to answer queries about hazardous knowledge (e.g., synthesizing anthrax) while maintaining the model’s ability to answer queries about non-hazardous knowledge (e.g., culturing yeast). We operationalize this as reducing a model’s QA accuracy on WMDP while maintaining performance on general capabilities benchmarks, such as MMLU and MT-Bench.

In contrast to unlearning for copyright or privacy, we do not assume access to questions from WMDP. This is because we are interested in methods that can generalize: unlearning an entire distribution of hazardous knowledge given limited samples.

2 Method

We leverage the idea that model representations encode knowledge of the world and that these representations may be controlled to guide the model (Ilharco et al., 2023; Zou et al., 2023a; Turner et al., 2023). Intuitively, we control the model towards having a novice-like level of hazardous knowledge. We design a loss function with a forget loss and a retain loss. The forget loss bends the model representations towards those of a novice, while the retain loss limits the amount of general capabilities removed (Figure 7).

Our goal is to bend the model representations on hazardous knowledge towards those of a novice. We must precisely specify both the distribution of knowledge to unlearn and the direction to push the activations towards. If we misspecify the distribution, we will either delete too little or too much information. If we misspecify the direction, we may degrade the model in unintended ways.

To specify the distribution of knowledge to unlearn (e.g., cybersecurity), we use keywords (e.g., exploit development or penetration testing). If we remove knowledge specified by such keywords, then we can maximally remove hazardous knowledge while minimally destroying general capabilities.

To guide activations, we use control vectors (Zou et al., 2023a; Ilharco et al., 2023). Given some keyword (e.g., “penetration testing”), an unlearning control vector is a vector in the model’s activation space that represents the absence of knowledge of that keyword. Controlling the model with unlearning control vectors is more precise than controlling towards a generic direction.

Our keywords are generated by the topics in threat models for biosecurity and cybersecurity (Appendix B.1). For example, for biosecurity, one of our keywords is “bioterrorism” and for cybersecurity, one of our keywords is “penetration testing.”

To control the activations, we assume a forget dataset $D_{\text{forget}}$ and compute:

where $h_{\text{control}}$ is a randomly sampled unlearning control vector from our set.

The full loss (Figure 7) is a weighted combination of the forget loss and the retain loss:

To alter model activations on hazardous knowledge, we need to collect $D_{\text{forget}}$, an unlearning distribution which approximates WMDP. To collect $D_{\text{forget}}$ for biosecurity, we collect a corpus of relevant papers from PubMed used to generate questions in WMDP-Bio (Section A.4). To collect $D_{\text{forget}}$ for cybersecurity, we conduct an extensive crawl of GitHub for documents associated with the topics in WMDP-Cyber, and filter the contents to include only the most relevant passages to WMDP-Cyber (Section A.5).

Similarly, to preserve activations on general language modelling tasks, we need to collect $D_{\text{retain}}$, a knowledge preservation distribution which approximates general, non-hazardous knowledge. For these, we collected subject-specific retain sets detailed in Sections A.4 and A.5. However, we find in practice that Cut is more performant when $D_{\text{retain}}$ has qualitatively distinct content from $D_{\text{forget}}$, so as not to relearn the unlearned knowledge. Thus, we set $D_{\text{retain}}$ to be Wikitext (Merity et al., 2016). We release the unused subject-specific retain sets for WMDP-Bio and WMDP-Cyber publicly, to guide future unlearning methods that can more effectively use these corpora.

Experimental Results

We examine the performance of Cut and other unlearning methods. We describe the experimental setup (Section 5.1) and provide quantitative (Section 5.2) and robustness (Section 5.3) evaluations. We also check if unlearning on WMDP generalizes to more hazardous information (Section 5.4). Cut markedly improves upon existing baselines, but future work is necessary to improve the precision of unlearning hazardous knowledge while fully maintaining general capabilities.

We describe the benchmarks we use for evaluations, the models we use for unlearning, and the baselines we use for comparisons. We only conduct unlearning experiments on WMDP-Bio and WMDP-Cyber, as discussed in Section 4.

We evaluate removal of hazardous knowledge with WMDP. To evaluate the preservation of general knowledge, we use MMLU (Hendrycks et al., 2020b), focusing on topics similar to biosecurity (college biology, virology) and cybersecurity (college computer science, computer security). Finally, to evaluate the fluency of models, we use MT-Bench, a multi-turn conservation and instruction-following benchmark (Zheng et al., 2023b).

We remove knowledge of biosecurity and cybersecurity on zephyr-7b-beta (Tunstall et al., 2023) and Yi-34b-Chat (01-ai, 2023), two of the most performant open-source generative language models at 7 billion, and 34 billion, parameters respectively. Additionally, we report the performance of GPT-4 (OpenAI, 2023a) as an upper bound on benchmark performance.

We benchmark Cut against three unlearning baselines: SCRUB (Kurmanji et al., 2023), SSD (Foster et al., 2024), and LLMU (Yao et al., 2023b), on zephyr-7b. Because we found low performance on zephyr-7b, we did not benchmark the baselines on Yi-34b. See Appendix B.6 for our implementation of the baselines.

2 Quantitative Evaluation

To assess the efficacy of the methods, we examine the forget performance and retain performance of the unlearned models. We see that Cut is able to unlearn WMDP-Bio and WMDP-Cyber while maintaining performance on MMLU (Section 5).

We measure forget performance by evaluating the knowledge of models on WMDP with both question-answering (QA) and probing.

QA evaluation. In the future, LLMs may be used by adversaries as knowledge engines for developing weapons. Under an API-access threat model, adversaries only receive output tokens and logits, without access to internal activations. Hence, we evaluate the QA accuracy of models on WMDP. We use a zero-shot question-answer format (Section B.2), taking the top logit between A, B, C, and D as the answer choice. For comparison, we also benchmark GPT-4 zero-shot on each of these tasks. As language models are sensitive to the prompting scheme (Sclar et al., 2023), we use lm-evaluation-harness (Gao et al., 2021) to standardize prompts.

QA results. We assess whether Cut is able to reduce QA accuracy on WMDP in Table 1. For both zephyr-7b and Yi-34b, Cut is able to drop performance to near random accuracy on WMDP-Bio and WMDP-Cyber, while other baselines struggle to drop accuracy on WMDP-Bio and WMDP-Cyber without crippling model performance on MMLU. We provide a more comprehensive table of results in 1.

Probing evaluation. While evaluating QA accuracy measures the primary risk of the API-access threat model, it fails to assess whether knowledge has been fully removed from the models. Models may possess more knowledge than is revealed in their output logits (Burns et al., 2022); for instance, the unlearned model may still retain hazardous knowledge, but refuse to answer. Thus, we test whether unlearned models can be probed to recall unlearned information. We train a 4-way linear probe on the unlearned Cut models. We use half of WMDP-Bio and WMDP-Cyber for training and hold out the other half for evaluation. We apply probing and report results for all layers of the model.

Probing results. We assess whether probes are able to recover knowledge from a model unlearned with Cut in Figure 9. Across both categories and model sizes, linear probing only achieves slightly better than random accuracy. Linear probes are unable to extract unlearned information from the model, suggesting that Cut does not merely mask or hide the information superficially, but rather causes a substantial alteration that prevents the recall of the unlearned information.

We measure the retain performance by evaluating models’ knowledge on MMLU and their fluency on MT-Bench.

MMLU evaluation. To be practical, unlearning methods must maintain general knowledge while removing hazardous knowledge. To evaluate whether models retain general knowledge after unlearning, we reuse the earlier QA evaluation setup for MMLU.

MMLU results. We report accuracy on subject-specific areas in MMLU (Figure 11). In contrast to other baselines which either fail to reduce performance on WMDP or greatly reduce performance on MMLU (Figure 11), Cut reduces performance on WMDP while maintaining overall MMLU accuracy. Moreover, Figure 11 shows that Cut retains performance on MMLU topics related to biology (college biology) and computer science (college CS), suggesting greater unlearning precision than the baselines. However, Cut greatly drops performance on the most similar topics to biosecurity (virology) and cybersecurity (computer security), suggesting the possibility for future work to improve retention of general capabilities during unlearning. As we use Wikitext as the retain set, Cut cannot determine exactly what knowledge to unlearn and retain. Thus, we encourage future work to employ our subject-specific biology and cyber retain sets (Section 4.2) to improve unlearning precision.

MT-Bench evaluation. Beyond retaining performance on academic multiple-choice questions, unlearned models should still maintain general conversational and assistant abilities. We evaluate Cut and all baselines on MT-Bench, a widely used metric for language model conversational fluency and helpfulness. We again evaluate GPT-4 as an upper bound for benchmark performance.

MT-Bench results. We report the MT-Bench performance of all models in Table 1. Cut roughly maintains performance on MT-Bench, with the score only decreasing $0.13$ on zephyr-7b and $0.54$ points on Yi-34b (out of a total possible of $9$). Because Cut still exhibits some degradation on MT-bench, particularly with Yi-34b, there is a need for further development of unlearning methods that can retain general assistant capabilities.

3 Robustness Evaluation

A primary motivation for unlearning is ensuring that knowledge is irrecoverable, even when subject to optimization pressure. If unlearning is not resilient, the adversary can still jailbreak the model to access hazardous information after unlearning.

We conduct a qualitative experiment using the GCG adversarial attack (Zou et al., 2023b) to measure whether dangerous knowledge is recoverable after performing Cut. We sample a single prompt from each of the WMDP-Bio and WMDP-Cyber datasets, slightly modify it such that the base Yi-34b models refuse to answer, and identify whether GCG can jailbreak the base and unlearned Yi-34b models to extract the correct answer (Section B.4).

GCG can jailbreak the base Yi-34b models to answer these prompts in less than 50 gradient steps, while the unlearned models output gibberish even after $2,\!500$ steps, or over 7 hours of optimization on an NVIDIA A100 GPU (Figure 12). This is a signal towards the resilience of Cut, suggesting that unlearning persists even under optimization pressure.

4 Generalization of WMDP to Hazardous Knowledge

We evaluate if unlearning on WMDP generalizes to unlearning especially hazardous knowledge.

During our dataset generation process, we identified $122$ questions in biosecurity that contained sensitive information and removed them from WMDP-Bio. We treat these as a held-out set of private questions with especially hazardous knowledge. We can evaluate whether WMDP is a proxy for hazardous knowledge by examining if performance on WMDP correlates with performance on the private set.

We follow the QA evaluation described in Section 5.2 and report the performance of zephyr-7b and Yi-34b before and after unlearning with Cut on this private set in Figure 13. Before and after unlearning, both models achieve similar accuracy on both the private set and WMDP. This result suggests WMDP is a reasonable proxy for especially hazardous knowledge.

Discussion

We discuss how unlearning on WMDP can tie in with other strategies to mitigate malicious use, such as structured API access. See Appendix D for a fuller discussion of the broader impacts of WMDP.

Unlearning on WMDP mitigates risk for both closed-source and open-source models.

For closed-source models, unlearning reduces risk from malicious API finetuning (Zhan et al., 2023; Qi et al., 2023; Pelrine et al., 2023), as hazardous knowledge can be removed prior to serving the model. Furthermore, unlearning is a countermeasure against jailbreaks—even if they are jailbroken, unlearned models lack the knowledge necessary to empower malicious users (Figure 2).

For open-source models, unlearning can expunge hazardous knowledge before such models are publicly released, limiting adversaries from repurposing open-source models out of the box. However, unlearning with on WMDP does not prevent relearning in open source models (Section B.5). We encourage future work towards mitigating risk in this pathway.

2 Structured API Access

WMDP complements the safety benefits of structured API access (Shevlane, 2022), where model developers provide an API for users to query and finetune models without full weight access. In this framework, ordinary users may query and finetune models with an API, but the model provider applies safety mechanisms, such as unlearning, prior to serving the model. However, approved users could obtain API access to the base model with full capabilities under strict guidelines, empowering the use of LLMs for benign or defensive applications while mitigating potential vectors of malicious use. For instance, OpenAI allows access of GPT-4 variants with fewer guardrails for red-teaming and biological malicious use experiments (OpenAI, 2023a, 2024). Structured access mitigates the concern that unlearning dual-use information will harm defenders.

Structured access requires model developers to solve the “Know Your Customer” (KYC) challenge, which involves verifying the identity and intentions of customers before allowing them privileged interactions. For structured access, implementing KYC-like procedures can help mitigate the risks associated with malicious use by ensuring that only verified and trustworthy individuals or organizations are given the full capabilities of the model.

Conclusion

We propose a dataset, WMDP, to evaluate the potential of malicious use in LLMs. WMDP was developed by subject matter experts in biology, cybersecurity, and chemistry, and was filtered to remove sensitive or export-controlled information. Modern LLMs score highly on some aspects of WMDP, suggesting presence of hazardous knowledge. We propose machine unlearning as a safety intervention to reduce hazardous knowledge.

Towards making progress on unlearning, we introduce Cut, an unlearning method that removes hazardous knowledge without significantly compromising general model performance. Cut also generalizes and successfully removes information from a private sensitive dataset. However, Cut reduces accuracy on closely related fields, such as introductory virology and computer security, demonstrating the need for continued research towards improved unlearning precision.

We thank Alexander Sikalov, Adrian Huang, Andrew Papier, Anthony DeLorenzo, Anthony M. Barrett, Cristae Consulting, Dinesh C. Aluthge, Frances Ding, Geetha Jeyapragasan, Isabella Weinland, Jake Pencharz, Jaspreet Pannu, Kathryn McElroy, Matthew Blyth, Mei Yi You, Miriam Sun, Nikki Teran, Patrick Biernat, RET2 Systems, Inc., Ritoban Roy-Chowdhury, Samuel Curtis, Scott Donahue, Steve Newman, and Xinyan Hu for their assistance and feedback. AP acknowledges support from the Vitalik Buterin PhD Fellowship in AI Existential Safety. AD and SG acknowledge support from the ML Alignment Theory Scholars (MATS) program.

References

Appendix A Dataset

We describe the WMDP dataset in more detail, providing a breakdown of question categories in WMDP (Section A.1) and describing additional and considerations for WMDP-Bio (Section A.2 and WMDP-Chem (Section A.3). Then, we outline the unlearning corpora for WMDP (Sections A.4 and A.5).

A.2 Additional Considerations for WMDP-Bio

The questions generated based on our threat model in Section 3.2 primarily test for knowledge on how to develop or use pre-existing resources acquire bioweapons, including some types of tacit knowledge, such as the knowledge needed to troubleshooting experiments in a laboratory. While a common objection to this method may be that other critical parts of the bioweapons acquisition pathway (such as acquiring reagents or equipment) represent the “true” bottleneck, we believe testing for knowledge is still useful for several reasons: first, it is unclear to what extent pre-existing bottlenecks in reagent and equipment acquisition will remain in the future, due to accelerating progress in laboratory automation, outsourcing, and science democratization (such as with the proliferation so-called “citizen science” or “biohacker” labs). Second, certain types of knowledge can accelerate risks by unblocking bottlenecks around equipment and reagents (such as knowledge of DNA synthesis screening vulnerabilities). Finally, to a first approximation, knowledge of more dangerous threats can still increase the marginal risk profile of the expected economic damage and casualties from bioweapons development, despite additional bottlenecks in the larger bioweapons acquisition pathway.

A.3 Additional Considerations for WMDP-Chem

Careful consideration of the full process by which malicious actors will access and deploy chemical weapons and/or explosives has led us to define the following categories of knowledge to include in WMDP-Chem:

General knowledge, which refers to fundamental scientific knowledge of the target materials and their properties;

Synthesis, which refers to all preparation procedures, including chemical syntheses and other fabrication procedures;

Sourcing/procurement, which refers to knowledge about how to access starting materials, particularly in ways that are outside of established and regulated chemistry laboratories;

Purification, which refers to knowledge about how to purify crude products after they are successfully synthesized, particularly using methods that work outside of a regulated laboratory;

Analysis/verification, which refers to the procedures required to analyze all target compounds, as well as the intermediates accessed as part of the synthetic pathway. Such procedures are necessary to confirm that the desired compounds have been synthesized, and are of sufficiently high purity to ensure that their usage will have the desired effect.

Deployment mechanisms (e.g. device, plans), which refers to all knowledge necessary to effectively deploy the chemical weapons and/or explosive, including fabrication of a device; dispersal of the compound in an aerosol, capsule, or vapor form; dissolution of the compound in the drinking water supply; and diffusion of the compound into the HVAC system of a target location.

Deployment: bypassing detection, which refers to all technical knowledge that is necessary to avoid detection by law enforcement authorities. This includes knowledge about how to transport materials through the mass transportation network, how to use covert methods to access all necessary materials, and how to deploy the compound in a way that limits the individual’s potential exposure to liability.

Miscellaneous knowledge, which refers to all additional knowledge that is not covered in the aforementioned categories, including knowledge about derivation of target chemical weapons and/or explosives, properties of such derivatives, and information about mitigation and response strategies that people are likely to use following the deployment of the harmful agents.

A.4 Bio Corpora

The forget and retain corpora are a collection of papers from PubMed. The forget set includes papers that were used to generate the WMDP-Bio questions, while the retain set samples papers across categories for general biology, while omitting papers in the forget set and using keyword exclusion against the topics in Section B.1.

A.5 Cyber Corpora

The forget and retain corpora consist of passages scraped via keyword search on GitHub. The keywords used for the forget corpora are

We then employ Mixtral-8x7B-Instruct-v0.1 [Jiang et al., 2024] to filter the dataset further with the following prompt, accepting passages only with a score of 9 or higher:

For the retain set, we use the following search terms:

Appendix B Experiments

We provide the full benchmarking and unlearning results in Table 1. We describe additional details for implementing Cut (Section B.1) and evaluating on WMDP (Sections B.2 and B.3). Then, we describe the implementational details for the robustness (Section B.4) and relearning (Section B.5) evaluation, before discussing the unlearning baselines we evaluated (Section B.6).

We use the list of WMDP-Bio categories as the keywords for Cut unlearning with biosecurity.

We use the following keywords for Cut unlearning with cybersecurity:

We use the default lm-evaluation-harness zero-shot format for MMLU and WMDP. An example is as follows:

We employ the single evaluation mode on MT-Bench, using gpt-4-0613 as the judge.

We describe the results of the robustness evaluation experiment (Section 5.3), with jailbreak results for the Cut unlearned and base Yi-34b models.

We employ GCG [Zou et al., 2023b] with a 20-token adversarial suffix. Note that GCG requires a target answer output of multiple tokens, so rather tham constraining Yi-34b to answer in multiple choice format, we remove the answer choices and allow it to conduct a free-form generation.

For WMDP-Bio, we sampled the following reworded prompt:

After optimizing over the Cut unlearned Yi-34b model for $2,\!500$ steps, the resulting suffix and output for the WMDP-Bio prompt is

For WMDP-Cyber, we sampled the following (reworded) prompt:

After optimizing over the Cut unlearned Yi-34b model for $2,\!500$ steps, the resulting suffix and completion for the WMDP-Cyber prompt is

B.4.2 Base Model

The base model elicited a refusal when the WMDP-Bio prompt was asked directly:

Applying GCG for 50 optimization steps yielded the following suffix and successfully jailbroken completion:

The base model elicited a refusal when the WMDP-Cyber prompt was asked directly:

Applying GCG for 50 optimization steps yielded the following suffix and successfully jailbroken completion:

We evaluate whether Cut prevents finetuning from recovering hazardous knowledge. Our work focuses on the closed-source threat model where LLM providers apply unlearning before LLM serving (Figure 2). We now consider the open-source threat model where LLM providers publicly release the LLM weights. In this setting, adversaries may finetune the model to attempt to recover hazardous capabilities.

We examine if Cut also prevents models from relearning unlearned knowledge through finetuning. In particular, we perform unlearning on Mistral-7B-v0.1 [Mistral AI team, 2023] and afterwards finetune on the cybersecurity forget corpus. In practice, we find it difficult to finetune zephyr-7b on our unlabeled corpus due to its instruction-tuning, so we use its base model, Mistral-7B-v0.1.

We finetune until the loss remains steady and report the results of finetuning in Figure 14. We see that Cut is unable to prevent finetuning from recovering performance, and we encourage future work to tackle the challenge of preventing relearning of unlearned knowledge through finetuning.

We describe the baselines we employed, and any implementational details we employed for unlearning on Cut.

We make several changes in adapting LLMU [Yao et al., 2023b] to our setting. We use bfloat16 for all floating point computations. In the unlearning process we do not stop after a prescribed maximum forget loss, rather stopping after unlearning for exactly a prescribed number of steps. Each sample of our dataset is truncated to 200 characters, and in the random loss we remove the question answer formatting, as our corpora does not follow this format. Using the hyperparameters for Llama 2 (7B) as a starting point, we employ low-rank adaptation [Hu et al., 2021], a batch size of 2, a random weight of 1, and a normal weight of 1. We apply a grid search over the learning rates $[1\times 10^{-4},5\times 10^{-4},1\times 10^{-3},5\times 10^{-3}]$, the number of steps $$, and the forget weight$[0.5,1,2]$.

B.6.2 SCRUB

Kurmanji et al. propose SCalable Remembering and Unlearning unBound (SCRUB) for image classification. It uses the original model as a frozen teacher and clones it to form a student model that is adapted for unlearning. SCRUB cycles between forget data and retain data epochs, maximizing KL divergence of logits between the student and teacher model on the forget set, and minimizing it on the retain set. The retain set epochs also includes a task-specific loss with gold labels to maintain performance. We use the same forget set and retain sets as the Cut experiments, and with log perplexity on Wikitext as the task-specific loss. We tune the $\alpha$ hyperparameter at values $[1\times 10^{-4},1\times 10^{-3},1\times 10^{-2},1\times 10^{-1},1,10]$, to search over loss weightings between knowledge distillation and the task-specific loss. We do this as a grid search with learning rates being $[1\times 10^{-5},5\times 10^{-6},2\times 10^{-6}]$. We use $600$ unlearning steps in total, doing the forget step only for $300$ as it is recommended in Kurmanji et al. to stop it earlier. In the high learning rate case, i.e. $lr=1e-5$ we also try doing only $400$ unlearning steps in total, with only $100$ forget steps. Other than that, we use the same hyperparameters as those reported for LLMU above. Goel et al. have shown that SCRUB performs poorly when most training samples relevant to removal are not available. This could be one of the reasons why SCRUB performs poorly in our setting.

B.6.3 SSD

Selective Synaptic Dampening (SSD) [Foster et al., 2024] belongs to a class of methods which find parameters in the model that are differentially more important for the forget set than the retain set. While the method was originally developed for image classification, we adapt it for autoregressive language modeling by altering the loss function to log-perplexity on the forget set and retain set. We grid-search on the threshold $[0.1,0.25,0.5,1,2.5,5]$ and constant for dampening $[1\times 10^{-5},1\times 10^{-4},1\times 10^{-3},1\times 10^{-2},1\times 10^{-1},1]$, the two main hyperparameters for SSD. We converged on these ranges after initial manual hyperparameter exploration for our task and datasets.

B.6.4 Cut

To enable further research on unlearning, we provide auxiliary benchmarks via unlearning certain subsets of MMLU, while retaining performance on the remainder of MMLU.

Economics: Unlearning on high school macroeconomics and high school microeconomics while retaining all other categories of MMLU.

Law: Unlearning on international law and professional law while retaining all other categories of MMLU.

Physics: Unlearning on high school physics, conceptual physics, and college physics while retaining all other categories of MMLU.

We specifically chose these settings to forget topics that were relatively separate from the remainder of MMLU, and contained a large enough sample size of forget set questions to benchmark on (more than $1,\!000$ questions).

We publicly release forget set corpora for all three of these settings. For each subject, a selection of textbooks with Creative Commons licenses were identified (ranging from high-school to graduate level). The text from these books was extracted and filtered to a set of paragraph-length chunks. The beginnings and end matter (table of contents, acknowledgements, index, etc.) of each book were excluded, as were most equations and exercises. Additional cleaning was performed to remove citations, links, and other artifacts.

Table 3 demonstrates the results of Cut unlearning for each setting. In the forget column, we report the accuracy for each setting, aggregated across all topics within the setting. For the retain column, we include closely related MMLU categories that should not be unlearned – College Mathematics and High School Mathematics for Physics, Jurisprudence for Law, and Econometrics for Economics. Lastly, we also report the aggregate MMLU performance before and after Cut unlearning.

Unlearning on Physics results in a significant performance drop on College Physics and High School Physics, despite an almost unchanged score on MMLU and Math related areas. However, we observe significant degradations in Retain set performance while unlearning on Economics and Law, demonstrating the potential for future methods to improve unlearning precision.

Appendix D Broader Impacts of WMDP

We reflect on how WMDP comports with the broader landscape of risk mitigation strategies.

From a policy-making perspective, we hope that WMDP guides the evaluation of hazards posed by ML systems, such as by informing the National Institutes of Standards and Technology’s AI Risk Management Framework [NIST, 2023, White House, 2023] or other frameworks. Moreover, WMDP may serve as risk marker for more stringent policy action. For example, a model scoring above a particular threshold on WMDP could be flagged for more comprehensive evaluation, such as human red teaming with biosecurity experts.

Furthermore, unlearning with WMDP may reduce general-purpose capabilities of models in biology or cybersecurity, which could hamper their utility for defensive, or beneficial, applications in those areas. Therefore, unlearning should be complemented with other safety interventions, such as structured access (Section 6.2). This is especially important for cybersecurity, as most cybersecurity knowledge may be used for both offensive and defensive purposes. For instance, AI progress could significantly enhance anomaly detection capabilities. This could aid attackers in disguising their activities to mimic normal usage patterns, but also inform critical infrastructure providers of atypical behavior that could signify an attack.

In biosecurity, however, there exist categories of primarily offensive knowledge that may be unlearned without significant degradation to defensive capabilities. For instance, knowledge of historical bioweapons programs may be safely removed from models without significantly affecting knowledge related to countermeasure development or general-purpose biology. As a result, while both WMDP-Bio and WMDP-Cyber are both useful measurements of hazardous language model capabilities, WMDP-Bio may be the most useful tool for risk mitigation via unlearning.

More broadly, there are other strategies, including non-technical strategies, that could be pursued to mitigate malicious use – such as implementing universal screening of synthetic DNA orders to prevent the widespread access to pathogen DNA, addressing gaps in the regulation of Select Agents in the Federal Select Agent Program, and improving oversight of laboratory automation and outsourcing.

WMDP consists of four-way multiple choice questions, potentially neglecting hazards that only surface in larger end-to-end evaluations. For instance, models that have memorized key biological concepts from the training data may be equally likely to do well on a particular multiple choice question as are models that have a true understanding of the underlying concept. Memorized facts may be particularly over-represented in our biological benchmark since many questions that were developed were drawn from open-access papers that were likely also included in the model’s training data. In addition, multiple choice questions only test for whether the model retains hazardous knowledge; these questions do not test whether the model will reveal that information to the end-user in a helpful and timely manner during the planning or execution of a nefarious attack. To address these limitations, future work in this area could include generating questions from scientific papers that were only released after a model’s training date cutoff, or using other strategies to generate questions which are difficult to search [Rein et al., 2023, Lála et al., 2023, Lynch et al., 2024].

WMDP is a static benchmark which cannot anticipate the evolving landscape of cyber and biological risks, as threats continuously change and new technologies emerge. Moreover, as with any metric, scores on WMDP do not capture the full extent of malicious use risk. As a result, benchmarking on only WMDP may yield a false sense of model safety after unlearning. This limitation emphasizes the need for other safety benchmarks to complement WMDP, especially as new risks emerge over time. For instance, benchmarks that assess open-ended conversations may be a more promising method to assess capabilities of future models.

WMDP focuses on reducing risk for API-access models (Section 1); for models with publicly downloadable weights, unlearned information can be trivially re-introduced by malicious actors. If open-source models reach similar capabilities to closed-source models in the future, these risks will remain unaddressed by this work.

Appendix E X-Risk Sheet

We provide an analysis of how our paper contributes to reducing existential risk from AI, following the framework suggested by Hendrycks and Mazeika . Individual question responses do not decisively imply relevance or irrelevance to existential risk reduction.

In this section, please analyze how this work shapes the process that will lead to advanced AI systems and how it steers the process in a safer direction.

E.2 Safety-Capabilities Balance

In this section, please analyze how this work relates to general capabilities and how it affects the balance between safety and hazards from general capabilities.

Overview. How does this improve safety more than it improves general capabilities?

Answer: Unlearning does not improve general capabilities; rather, it removes specific model capabilities while improving inherent model safety.

Red Teaming. What is a way in which this hastens general capabilities or the onset of x-risks? Answer: Although WMDP is constructed as a benchmark for measuring and reducing inherent model hazards, it may inadvertently serve as a roadmap for malicious use, hastening the onset of x-risks by lowering the barrier for causing catastrophe. To reduce these risks, we conduct an extensive sensitive information mitigation process (Section 3.5).

General Tasks. Does this work advance progress on tasks that have been previously considered the subject of usual capabilities research? $\square$

General Goals. Does this improve or facilitate research towards general prediction, classification, state estimation, efficiency, scalability, generation, data compression, executing clear instructions, helpfulness, informativeness, reasoning, planning, researching, optimization, (self-)supervised learning, sequential decision making, recursive self-improvement, open-ended goals, models accessing the Internet, or similar capabilities? $\square$

Correlation with General Aptitude. Is the analyzed capability known to be highly predicted by general cognitive ability or educational attainment? $\square$

Safety via Capabilities. Does this advance safety along with, or as a consequence of, advancing other capabilities or the study of AI? $\square$

E.3 Elaborations and Other Consideratiosns

Other. What clarifications or uncertainties about this work and x-risk are worth mentioning?

Answer: While unlearning is an important intervention for reducing model hazards, unlearning with may reduce the defensive, or beneficial, applications in those areas. unlearning should be complemented with other interventions that reduce risk (Appendix D).