Our Frontier AI Grantmaking in the First Half of 2026

In the first half of 2026, the Founders Pledge Global Catastrophic Risks Fund has directed over $1M into projects addressing the risks of frontier AI. We did this for a simple reason: those risks are no longer hypothetical.

Anthropic's Claude Mythos, the most advanced frontier AI system, can breach computer networks autonomously, identifying and exploiting security vulnerabilities faster and more reliably than most human specialists. We are likely entering a period where AI is used routinely in cyberwarfare.

And that is only the start: models capable of assisting in the development of catastrophic bioweapons, or of acting autonomously in ways that could undermine human oversight, are predicted to follow over the next few years. But governments and philanthropists are also moving quickly.

The EU is furthest ahead. On August 2, 2026, the EU AI Office will become the first regulator with binding enforcement authority over frontier AI. The US and UK safety institutes, CAISI and AISI, do excellent work testing models before deployment for dangerous properties, but operate voluntarily.

However, even the United States is now acting. On June 2, 2026, President Trump signed an executive order directing the government to evaluate frontier models for cyber capabilities before release, with developers granting agencies up to 30 days of early access. The underlying argument is settled: sufficiently capable models force a government response whatever the politics. The only open question is whether that response is adequate.

All of these agencies face similar problems: they are small teams that need to rapidly scale up and develop the techniques needed to track the various risks posed by frontier AI. This requires testing for the right risks, not just the obvious ones. It requires results that can be compared across models and labs, and it requires learning from previous rounds of testing rather than starting from scratch each time.

Without all of that, mandates produce the appearance of safety rather than the substance. This is where philanthropists can help: CAISI, AISI and the EU AI Office need to subcontract or rely on outside expertise simply to fulfil their basic mandates.

In the first half of 2026 we made five grants across two gaps. The first is capacity: regulators have the authority to test frontier models but not yet the means to test for the right risks, compare results, or build on what they learn. The second is durability: the institutions meant to use those evaluations can be dismantled or left to duplicate each other's work. The grants below address each in turn.

Broadening and improving independent AI evaluation

Frontier models are already surprisingly adept at manipulating people. When prompted, frontier models will often comply with requests to persuade a user toward harmful ends, including glorifying terrorism or endorsing political violence, and can be surprisingly persuasive, even today. In the future, when people rely on much more intelligent and autonomous systems to make decisions, this could become a serious source of risk, especially if those AI systems can also assist with causing catastrophic harm.

Despite this concern, Google is currently the only major lab to track persuasion/manipulation as a formal capability in its published safety framework. The EU AI Office has a mandate to enforce the safety standards covering these risks, but lacks the capacity to do so. Private nonprofits have stepped up to help them fulfill the mandate, but they need near-term funding before the enforcement deadline arrives.

FAR.AI Integrity Team ($500,000). FAR.AI's Attempt to Persuade Eval revealed that Google's Gemini would manipulate users into endorsing mass violence while refusing the same request made directly; after the disclosure, Google built its own manipulation evaluations and cut Gemini's compliance on those topics by over 50 percentage points. The grant funds the evaluations FAR.AI is now developing for the EU AI Office's consortium on manipulation, the work needed to make the AI Act's standards enforceable before the August deadline.

Apart Research ($200,000). Apart is running a six-month task force conducting manipulation evaluations for the EU AI Office before its August enforcement deadline. This grant funds a sprint team producing the metrics the Office needs to act before its new powers take effect. The results of these and other evaluations also need to be usable. Right now, with the frantic pace of progress and research, AI evaluation results are scattered across system cards, papers, and leaderboards with no common format. The same model tested on the same benchmark can produce different scores because testing conditions are not recorded. Researchers and policymakers cannot build on findings they cannot compare, and this is a general problem in a field moving this quickly.

EvalEval Coalition ($63,000). The EvalEval Coalition, a 400-member group hosted by Hugging Face, the University of Edinburgh, and EleutherAI, built a data standard that makes evaluation results comparable across models and labs. CAISI has adopted the schema. The grant funds Eval Cards, a public dashboard where policymakers and researchers can read, compare, and share these results for the first time.

These three grants fill urgent gaps in the world’s AI evaluation infrastructure, helping governments deal with the risks of frontier AI.

Coordinating AI governance

Building evaluation capacity is worthless if the institutions that use it get dismantled or trip over each other. The UK spent years and significant political capital building some of the strongest AI safety infrastructure in the world, including an AI Security Institute that tests frontier models before release. Political support for maintaining it has weakened under the current government, and what took years to build could be dismantled quickly.

Separately, when Claude Mythos was released, dozens of AI safety and governance organizations each had to track the incident, decide whether to engage, and produce a response independently, resulting in duplicated effort, missed windows, and contradictory public messaging. Both problems point to the same underlying weakness: the field has built real capacity but has not yet built the political and organizational infrastructure needed to protect and use it effectively.

Centre for Long-Term Resilience ($100,000). CLTR is a UK policy think tank that has worked directly with senior government officials on AI and biosecurity policy for several years. They are now building the communications capacity to defend the UK's position on AI safety: large-scale polling to understand where the public stands, a newsroom that has already placed coverage in The Times, the Financial Times, the Guardian, and BBC Radio, and coordinated messaging across pro-safety organizations. If the UK's safety infrastructure survives the current political climate, it becomes a model other countries can follow. If it doesn't, the cost of rebuilding it will be far higher than the cost of defending it now. Our grant funds the polling and communications capacity to defend that infrastructure during the narrow window when it is politically vulnerable, while defending it still costs far less than rebuilding it later.

Future Matters ($301,600). Future Matters, a Berlin-based nonprofit, runs standing coordination groups across roughly 130 organizations covering AI policy and biosecurity in both the US and internationally, so the field can share intelligence, align strategy, and respond together when something happens. Two international coordination groups were handed to Future Matters by other organizations in the network because they recognised the need for dedicated infrastructure and could not sustain it themselves. Our grant funds the standing coordination infrastructure no single organization can sustain alone, so the field can meet the next major model release as one effort rather than as 130 separate ones.

What comes next

AI is developing faster than the institutions designed to govern it. These five grants are designed to close specific gaps in how frontier AI systems are tested and how the organizations responsible for governing them work together. In later posts, we will lay out our longer-term thinking, including how advanced AI can itself be part of the solution to the risks it poses.