An independent 501(c)(3) building open-source safety evaluations that test frontier language models and AI agents for jailbreaks, prompt injection, deception, and corrigibility failures.
An independent 501(c)(3) building open-source safety evaluations that test frontier language models and AI agents for jailbreaks, prompt injection, deception, and corrigibility failures.
People
Updated 10/26/26 · By grantmaking.aiTeam Member
Org Details
Updated 10/26/26 · Provided by member · VerifiedSafe AI for Humanity Foundation is a Wyoming nonprofit (IRS 501(c)(3), approved April 2026) that builds independent, open-source evaluations for frontier language models and AI agents. We have no commercial affiliation with AI developers.
Our evaluation work covers adversarial instruction-following, jailbreak resistance, prompt injection, bias, hallucination, deception, and corrigibility. The harnesses, test suites, and results are public on GitHub (test-harnesses, ai4h-eval-lab, ai4h-test-suites, ai4h-evaluation-results).
We also publish incident-response plans for AI safety incidents and a public testing roadmap covering proprietary and open-weight models and the hardware needed to run repeatable evaluations in-house. We are early-stage and fully donor-supported.
Theory of Impact
Updated 10/26/26 · Provided by member · VerifiedCatastrophic risk from advanced AI increases when models are deployed with failure modes nobody outside the developer has measured. Independent, reproducible evaluations make those failure modes visible: (1) we build open test harnesses for behaviors that matter for loss-of-control and misuse risk, including deception, resistance to correction, prompt injection in agentic settings, and jailbreaks; (2) we run them against proprietary and open-weight models and publish results and methods openly; (3) developers, auditors, and policymakers can then reproduce, compare, and act on those results, raising the bar models must clear before deployment; (4) open-sourced harnesses, test, and tooling let other researchers extend the work rather than starting from scratch.
Projects– no linked projects
Updated 10/26/26 · By grantmaking.aiDiscussion
No comments yet. Be the first to share your thoughts.