Anthropic funds the development of independent AI model safety evaluations
Anthropic is launching a funding program for third-party safety evaluations, focused on the AI Safety Level, cybersecurity, CBRN risks, model autonomy, and testing infrastructure. Proposals can be submitted through an online form.
Anthropic has launched a funding program for safety evaluations developed by independent third-party organizations. The company says the goal is to strengthen the entire AI safety ecosystem, as demand for high-quality, safety-relevant evaluations exceeds supply, according to the company.
The program has three priority areas. The first is evaluations for the AI Safety Level (ASL) defined in Anthropic's Responsible Scaling Policy, specifically in the areas of cyber (models' ability to participate in cyberattacks comparable to advanced threats), chemical, biological, radiological, and nuclear risks (CBRN), model autonomy (including AI R&D tasks, self-replication, and weight exfiltration), other national security risks, social manipulation, and misalignment risks, where models, according to the company's research, may hide dangerous goals even after safety training. The second area is advanced capability metrics, for example expanding existing benchmarks such as GPQA with tens of thousands of new graduate-level questions and tasks. The third area is infrastructure, tools, and methods for creating the evaluations themselves.
The company invites organizations to submit proposals through an online form. The source text is incomplete, and a description of the specific funding terms, amounts, or timeline is missing from the available section — see the source article for details.
Why it matters
The program opens up a new funding path for organizations specializing in AI safety, for work they would otherwise have to cover from their own resources or grants. At the same time, it shows which types of risks (cyber, CBRN, autonomous model action, social manipulation) Anthropic considers a priority when assessing the safety of its models under the Responsible Scaling Policy, which could influence testing standards across the field.
Two audiences, two different impacts
What this means
For individuals
For individuals working in AI safety research or evaluation development, a concrete path to funding their own work is opening up; for the average user of AI tools, there is no direct impact.
For a business
For companies and organizations working on developing AI safety tests, this represents a new funding opportunity; for AI companies in general, it signals which areas (cyber, CBRN, model autonomy) are considered key for risk assessment under the Responsible Scaling Policy.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.