Fabraix

As a two-founder, just-launched company, Fabraix has already produced a working red-teaming agent with a published benchmark result, a fast initial pace for its stage.

Automated, environment-aware red-teaming would reduce the manual adversarial-testing burden security teams face when validating customer-facing AI agents, if it performs consistently at production scale.

As a company just entering its first YC batch with no disclosed funding or named customers, independently verifiable commercial momentum is essentially nonexistent at this stage.

Testing AI agents through active environment manipulation rather than prompt-only adversarial inputs reflects a more realistic and comprehensive red-teaming approach for agentic systems that can take real actions.

The 78% attack success rate on the named, independently published AgentHarm benchmark is a more concrete efficacy signal than pure vendor marketing claims, though it is self-reported and not independently reproduced by a third party here.

As customer-facing agentic AI deployment grows, adversarial testing against action-capable AI systems is likely to become a standard security requirement rather than a niche practice.