Fabraix

Fabraix is an adversarial verification platform for AI agents that probes agent systems for vulnerabilities and then enforces runtime guardrails to block them, built for developers and teams shipping AI agents to production. It combines offensive attack simulation with runtime defense through two products that work together: Nyx and Arx. Nyx autonomously stress-tests an agent using more than 1,000 adversarial strategies in a pure blackbox, multi-turn fashion, adapting to your system in real time to surface real exploits, reasoning gaps, instruction-following failures, and logic bugs that manual testing or static evals miss. Arx then translates those findings into runtime guardrails that actively block threats while the agent is live. Fabraix frames the work around its own research, including the Adversarial Cost to Exploit (ACE) framework, which reframes agent security from binary pass/fail into measurable, deployable economics, and it offers a hosted Playground for stress-testing agent defenses through adversarial play. The product is most relevant to engineering and security teams that need to find gaps in their AI agents before users (or attackers) do. Pricing is not published on the site.

Teams use Fabraix to adversarially test their AI agents for exploits and failure modes, then deploy runtime guardrails that block those attacks in production.