Mercor AI Safety Fund Grants

What we're looking for

We're open to proposals across the full range of safety interests. We're particularly interested in:

• Misalignment: deceptive alignment, goal misgeneralization, reward hacking, scheming, and situationally-aware failure modes

• Sandbox escape: containment failures, privilege escalation, tool misuse, and agents operating outside their intended scope

• Evaluation awareness: models detecting they are being tested and behaving differently under observation than in deployment

• Interpretability: understanding what models are actually doing internally, and whether that can be made legible to a human reviewer

• Oversight and control: scalable supervision, human-in-the-loop reliability, and what breaks when the system is more capable than its reviewer

• Red-teaming methodology: more robust systems for uncovering novel failures

If your work doesn't fit neatly into these, apply anyway. Strong proposals outside this list are welcome.