Center for AI Safety

The Center for AI Safety (CAIS, pronounced "case") is an American artificial intelligence safety research nonprofit headquartered in San Francisco, founded in 2022 by Dan Hendrycks. Hendrycks is a Berkeley computer-science PhD (under Jacob Steinhardt and Dawn Song; advisor of Transluce co-founder Steinhardt's later research group) whose academic-research output as a PhD student established several of the principal foundation-model evaluation benchmarks that have become industry standards: MMLU (Massive Multitask Language Understanding, 2020), MATH (competition mathematics benchmark, 2021), and HumanEval-adjacent code-evaluation contributions. CAIS's research mandate is oriented around AI safety with explicit existential-risk framing, technical AI safety research output, and substantial public-policy engagement. The May 2023 Statement on AI Risk — a single-sentence statement reading "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war" — was organized by CAIS and signed by approximately 350 AI researchers and executives including Sam Altman, Demis Hassabis, Geoffrey Hinton, and Yoshua Bengio, and was characterized in international coverage as a defining moment in the broader AI safety policy discourse. As of April 2026, CAIS is one of the principal independent AI safety research nonprofits globally alongside Apollo Research, METR, Conjecture, Transluce, Timaeus, MIRI, and the Berkeley Center for Human-Compatible AI.

At a glance

• Founded: 2022 in San Francisco by Dan Hendrycks. US 501(c)(3) nonprofit research organization.
• Status: Independent nonprofit research organization. Operates with a research-and-policy structure rather than a commercial company structure.
• Funding: AI safety philanthropic backing including Open Philanthropy, the Survival and Flourishing Fund, the Future of Life Institute, and adjacent AI-safety-focused funders. Specific cumulative funding figures are not comprehensively disclosed but industry coverage has reported multi-million-dollar annual operating budget.
• CEO / Lead: Dan Hendrycks, Director and Founder of CAIS. PhD computer science (UC Berkeley); developer of the MMLU, MATH, and adjacent foundation-model evaluation benchmarks; one of the more visible public AI safety voices through 2022 to 2026.
• Other notable leadership: Senior research staff across the technical AI safety research and policy-engagement programs. Hendrycks also serves as an advisor to xAI (Elon Musk's frontier AI lab) and other AI organizations, with the advisory roles shaping CAIS's adjacent industry engagement.
• Open weights: Yes, partial. CAIS produces open-source AI safety research and benchmark releases. The Weapons of Mass Destruction Proxy (WMDP) benchmark and adjacent dangerous-capability evaluation infrastructure are released open-source.
• Flagship outputs: MMLU (Massive Multitask Language Understanding benchmark, September 2020); MATH (competition-mathematics benchmark, March 2021); the May 2023 Statement on AI Risk; WMDP (Weapons of Mass Destruction Proxy benchmark for dangerous-capability evaluation, March 2024); active publication record on AI safety, dangerous-capability evaluation, and adjacent research areas; the SafeBench AI safety evaluation aggregator.

Origins

The Center for AI Safety was founded in 2022 in San Francisco by Dan Hendrycks, who had recently completed his PhD in computer science at UC Berkeley under the supervision of Jacob Steinhardt and Dawn Song. Hendrycks's academic-research output as a PhD student had been substantively consequential for the broader foundation-model evaluation landscape: the September 2020 release of MMLU (Massive Multitask Language Understanding) became one of the principal foundation-model evaluation benchmarks across academic and industry research, with subsequent MATH (competition mathematics) and HumanEval-adjacent code-evaluation contributions following in 2021. The benchmark-development work positioned Hendrycks within both the academic-AI-research and AI-safety-research communities by the time of the CAIS founding.

The 2022 founding period built CAIS as a research-and-policy nonprofit with an explicit existential-risk-focused mandate, distinguishing the organization from commercial AI safety alternatives (Anthropic safety research) and from AI safety research with different philosophical orientation (CHAI's preference-based alignment, MIRI's decision-theoretic alignment). Hendrycks's broader research output through 2022 to 2023 included foundational papers on robustness, distribution shift, AI safety methodology, and the broader question of how to evaluate increasingly capable AI systems.

The May 2023 Statement on AI Risk was the organization's most internationally visible early output. The statement was a deliberately concise single-sentence formulation: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." The statement was signed by approximately 350 AI researchers and executives at the time of release, including frontier-AI lab CEOs (Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, Dario Amodei of Anthropic), Turing Award laureates (Geoffrey Hinton, Yoshua Bengio), and adjacent senior AI researchers. The statement was characterized in international coverage as a defining moment in the broader AI safety policy discourse, with the signatory list functioning as a who's-who of the senior AI research community endorsing the existential-risk framing that AI safety advocates had previously argued was under-represented in mainstream discourse.

The 2023 to 2024 period saw substantive technical AI safety research output. The March 2024 release of the WMDP (Weapons of Mass Destruction Proxy) benchmark, with co-authorship across CAIS and external research collaborators, established a public benchmark for evaluating frontier AI models on dangerous-capability dimensions including biology, chemistry, and cybersecurity. The WMDP benchmark has been adopted by multiple AI safety evaluation programs including the UK AI Safety Institute and US AI Safety Institute (NIST) for pre-deployment frontier-model evaluation.

The 2024 to 2026 period has continued technical AI safety research output and policy engagement. Hendrycks's advisory engagement with xAI (Elon Musk's frontier AI lab) and adjacent industry organizations has shaped CAIS's industry-adjacent positioning. CAIS has continued to publish AI safety research, evaluation benchmarks, and policy submissions to UK, EU, US, and international AI safety regulatory processes.

Mission and strategy

The Center for AI Safety's stated mission is to reduce societal-scale risks from artificial intelligence. The strategy combines three threads. First, technical AI safety research output, including dangerous-capability evaluation benchmarks (MMLU, MATH, WMDP, SafeBench), AI safety methodology papers, and adjacent academic-research engagement. Second, public communication and policy engagement on AI risk, with the May 2023 Statement on AI Risk as the most prominent early example and continued submissions to UK, EU, US, and international AI safety regulatory processes. Third, the SafeBench AI safety evaluation aggregator and adjacent infrastructure that supports the broader AI safety research community.

The competitive premise is that AI safety requires both rigorous technical research and effective public communication, and that an independent nonprofit organization can pursue both threads in ways that commercial AI labs (with structural commercial-pressure-vs-safety conflicts) and academic research organizations (with longer publication cycles and less policy engagement) cannot fully match.

Models and products

• MMLU. Massive Multitask Language Understanding benchmark. Released September 2020 by Hendrycks and collaborators. The principal foundation-model knowledge-evaluation benchmark across academic and industry research.
• MATH. Competition mathematics benchmark. Released March 2021. Widely used for foundation-model mathematical reasoning evaluation.
• May 2023 Statement on AI Risk. Single-sentence statement organized by CAIS and signed by approximately 350 AI researchers and executives.
• WMDP. Weapons of Mass Destruction Proxy benchmark for dangerous-capability evaluation across biology, chemistry, and cybersecurity. Released March 2024.
• SafeBench. AI safety evaluation aggregator providing public-facing benchmarking infrastructure.
• Active publication record. AI safety, dangerous-capability evaluation, robustness, distribution-shift, and adjacent research papers.

Distribution channels are predominantly academic publication, open-source benchmark releases through GitHub, and policy submissions to UK, EU, US, and international AI safety regulatory bodies.

Benchmarks and standing

CAIS's evaluation framework focuses on the impact of its research outputs (with MMLU, MATH, and WMDP as widely adopted benchmarks), the visibility of its policy-engagement activities (with the May 2023 Statement on AI Risk as the most prominent example), and the broader influence on AI safety research and policy discourse.

Industry coverage has consistently characterized CAIS as one of the principal independent AI safety research organizations globally, with the Hendrycks-anchored research credibility, the substantial benchmark-and-evaluation infrastructure output, and the May 2023 Statement on AI Risk as principal validating data points.

Leadership

As of April 2026, CAIS's senior leadership includes:

• Dan Hendrycks, Director and Founder.
• Senior research staff across the technical AI safety research and policy-engagement programs.

Hendrycks also serves as an advisor to xAI and adjacent AI organizations, with the advisory roles shaping CAIS's industry-adjacent positioning.

Funding and backers

AI safety philanthropic backing including Open Philanthropy, the Survival and Flourishing Fund, the Future of Life Institute, and adjacent AI-safety-focused funders. Specific cumulative funding figures are not comprehensively disclosed but industry coverage has reported multi-million-dollar annual operating budget.

Industry position

CAIS occupies a distinctive position as one of the principal independent AI safety research nonprofits globally, with the substantial benchmark-and-evaluation infrastructure (MMLU, MATH, WMDP, SafeBench), the May 2023 Statement on AI Risk as the most prominent recent AI safety policy moment, and the Hendrycks-anchored research and public-engagement credibility. Industry coverage has consistently grouped CAIS with Apollo Research, METR, Conjecture, Transluce, Timaeus, MIRI, and the Berkeley Center for Human-Compatible AI as the principal independent AI safety research organizations of the post-2022 era.

The structural risks are two. First, the existential-risk framing that anchors CAIS's positioning has produced sustained engagement from a substantial AI safety community but has also produced criticism from researchers who argue that more immediate AI risks (bias, misuse, displacement) deserve comparable or greater research attention. Second, the Hendrycks-anchored leadership and visibility creates organizational dependency that key-person risk amplifies.

Competitive landscape

• Apollo Research, METR, Conjecture, Transluce, Timaeus, Goodfire. Independent AI safety research peer organizations.
• MIRI (Machine Intelligence Research Institute), CHAI (Berkeley Center for Human-Compatible AI). Earlier-cohort AI safety research organizations.
• UK AI Safety Institute, US AI Safety Institute (NIST). Government AI safety research bodies. Cooperation partners on benchmarks and evaluation methodology.
• Anthropic safety teams, Google DeepMind AGI safety, OpenAI Safety Systems. Frontier-AI-lab internal safety research organizations.
• xAI. Hendrycks advisory engagement; adjacent industry organization.
• Schmidt Sciences, Open Philanthropy, Future of Life Institute. AI safety philanthropic-funding peer organizations.

Outlook

• Continued technical AI safety research output and benchmark releases.
• Continued public-communication and policy engagement on AI risk through 2026 to 2027.
• Continued WMDP and SafeBench infrastructure development.
• The competitive dynamic with frontier-lab safety teams and other independent AI safety research organizations.
• Continued senior research-talent recruitment and retention.

Sources

• Center for AI Safety. Organization reference.
• Dan Hendrycks Wikipedia. Founder reference.
• Statement on AI Risk (May 2023). Public AI safety statement.
• WMDP benchmark. Dangerous-capability evaluation benchmark.
• MMLU paper. Foundation-model knowledge-evaluation benchmark.
• MATH paper. Competition-mathematics benchmark.