Project summary

We are building an independent, open-source benchmark that measures safety-critical behaviors in frontier large language models and tool-using AI agents, and releasing all code, prompts, and results publicly. We accept no funding from the AI developers whose models we test, so our results can serve as a neutral reference.

What are this project's goals? How will you achieve them?

Our goal is a trusted, reproducible safety benchmark covering adversarial instruction-following (jailbreaks and multi-turn manipulation), shutdown and scope compliance in tool-using agents, cross-modal safety consistency, harmful-content refusal across languages, and corrigibility in long-horizon agentic tasks. We will release v0.1 on GitHub, run evaluations across 8-12 frontier models, calibrate scoring rubrics against human-expert judgment, and publish a preprint alongside the full dataset and evaluation code.

How will this funding be used?

API compute: $9,000 for multi-turn, tool-calling evaluations across 8-12 frontier models. Researcher time: $4,500 for analysis, rubric calibration, publication, and codebase maintenance. Publication and dissemination: $1,500. Overhead is negligible; close to 100% is direct program spend, and all outputs are released under a permissive license (Apache 2.0 / CC-BY).

Who is on your team? What's your track record on similar projects?

The project is led by founder Aureliusz DeSimone, whose background in AI security research and infrastructure engineering maps directly onto adversarial testing and reproducible evaluation pipelines. Early evaluation harnesses and methodology are already published at ai-4-h.org. Safe AI for Humanity Foundation is an IRS-recognized 501(c)(3) public charity (EIN 41-4767005).

What are the most likely causes and outcomes if this project fails?

The main risks are model API changes or compute costs reducing coverage, or scoring rubrics proving difficult to calibrate reliably. Even in the worst case, the open framework and pilot results across a few models remain a reusable public good; partial funding still yields a working v0.1 and a public proof of concept that other researchers can extend.

How much money have you raised in the last 12 months, and from where?

Approximately $1,900 in donations to date. Applications to the Long-Term Future Fund and the Survival and Flourishing Fund are pending (submitted; no decision yet).