A finding is only as good as the setup that produced it. Our engineering work exists so that results can be repeated, compared, and checked by others.
Controlled environments that run the same experiment across models, versions, and defenses, so results can be compared.
Scored test sets and rubrics for behaviors that standard benchmarks miss, such as instruction drift and unreported actions.
Tooling that records what an agent was asked, what it reasoned, what it did, and what it said it did.
Logged adversarial attempts and outcomes, released where doing so does not create new risk.
Successive generations of automated attackers run against a single guarded-secret lab. Every probe is logged, along with the human solutions that beat them.
View on GitHub →Compares an agent session across four records: the request, the reasoning, the actions taken, and the final report. It flags where they disagree.
Tools that help people evaluate and defend AI systems are released under open licenses on GitHub. Tooling that would mainly make attacks easier stays private until patched.