Fabraix

Fabraix is an AI security research lab that builds adversarial testing tools for AI agents. The company takes an offensive security approach, operating on the principle that safety advances when you actively try to break things. Their open source tools let you probe AI systems for vulnerabilities, discover weaknesses through automated attacks, and then defend against what you find. If you're deploying AI agents and want to stress-test them before something goes wrong in production, this is the kind of tooling you need. The work draws from original research including ACE, which stands for Adversarial Cost to Exploit, a framework that translates security from binary pass-fail assessments into deployable economic models that quantify how much effort an attacker would need to compromise a system.

The core offensive product is Nyx, a fully autonomous AI agent that systematically probes other AI systems for security vulnerabilities. It runs multi-turn, blackbox attacks with real-time adaptation, meaning it doesn't need access to internals and adjusts its approach based on what it discovers. You provide a target and a budget, and Nyx generates detailed markdown audit reports on completion while tracking costs throughout the assessment. Coverage scales with compute rather than team size, so a small security team can run comprehensive adversarial tests that would otherwise require a dedicated red team. Nyx is released under an Apache 2.0 license and lives on GitHub as a TypeScript project you can run on your own infrastructure. The concurrent adversarial strategies mean it can test multiple attack vectors simultaneously rather than running them sequentially.

Alongside the offensive tooling, Fabraix offers Arx, a runtime defense system informed by what Nyx discovers. Arx validates agent actions before execution to detect goal deviation, logging events for observability and security analysis while managing agent session initialization with trace tracking. It's designed to block prompt injection, goal deviation, and memory poisoning in deployed agents. The idea is that the attacks you find during testing directly inform the defenses you deploy, creating a feedback loop between offense and defense. This mirrors how traditional security works in other domains, but applied specifically to the new attack surfaces that AI agents introduce. The session management includes trace tracking so you can reconstruct exactly what happened when an incident occurs.

The Playground at playground.fabraix.com is a live environment where you can stress-test AI agent defenses through adversarial play. It's useful for getting hands-on with the concepts before committing to a larger deployment. The playground itself is open source under MIT license with 71 stars and 11 forks on GitHub as of mid-2026, indicating active community interest. CrabTrap is their LLM-as-a-judge HTTP proxy for securing agents in production, a Go project forked from Brex's original implementation and released under MIT license. Having multiple tools under permissive licenses means you can mix and match based on where your system is in its lifecycle. The combination covers development testing, staging verification, and production monitoring.

Fabraix targets teams building AI agents that interact with external systems or handle sensitive operations. Financial services, healthcare, and anything touching infrastructure are obvious fits. But even consumer-facing AI products benefit from adversarial testing before launch. The question isn't whether your agent can be manipulated, it's how easily and what the consequences are. The documentation at docs.fabraix.com includes quickstart guides that let you begin either adversarial scanning or agent instrumentation within five minutes. Essential guides cover authentication, the Event data model, agent lifecycle integration, and foundational concepts. Community engagement happens through Discord for real-time discussion and troubleshooting.

The open source model means you can start using Nyx, Playground, and CrabTrap today without contacting sales or waiting for access. All four public repositories are actively maintained, with Playground and Nyx seeing updates through mid-2026. Pricing for any managed services or enterprise support isn't published on the site, but the core tools are genuinely free to run yourself. The trade-off is that you need the technical ability to deploy and operate the tools, which limits the audience to teams with engineering resources rather than business users looking for a turnkey solution.

What makes this approach different from manual pen testing is the systematic coverage. A human red teamer might find clever exploits, but Nyx runs exhaustive attack patterns that would take weeks to execute manually. The economic framing from ACE means you get quantified risk assessments rather than binary findings. You know not just that an attack is possible, but how much it would cost an adversary to execute, which helps prioritize fixes. For teams shipping AI agents into production, this is the gap between hoping nothing goes wrong and actually knowing what your exposure looks like.