Fabraix

What it does

Fabraix builds autonomous AI agents that adversarially test other AI agents for security vulnerabilities. Their flagship product, Nyx, is a black-box testing harness that probes AI systems across security, logic, and alignment. Nyx uses over 10,000 distinct attack strategies—including jailbreaks, prompt injections, exfiltration attempts, and reasoning exploits—in multi-turn, adaptive campaigns. It operates without any integration or internal access, simulating real-world attacks to surface failure modes that static benchmarks and manual audits miss.

Fabraix also offers an open-source Playground where researchers and security enthusiasts can run attack campaigns against their own systems and participate in monthly challenges to break live AI agents for rewards.

Who it is for

Fabraix targets any organization deploying customer-facing AI agents, including:

• Enterprise security teams needing continuous red-teaming for AI systems in production.
• AI startups building chatbots, coding agents, browser agents, voice assistants, or multi-agent systems.
• Financial services (e.g., trading agents, financial advisors) requiring compliance and safety testing.
• Healthcare (clinical copilots) needing to probe for PHI leakage, unsafe triage, or adversarial inputs.
• Researchers in AI safety and adversarial robustness, who can use the free tier and open-source Playground.

Fabraix claims to have found vulnerabilities in agents at dozens of Fortune 500 companies, though specific names are not disclosed on the website.

Why it matters

AI agents fail in ways traditional software does not—they are non-deterministic, vulnerable to prompt injection, and can be exploited across multiple turns. Traditional point-in-time audits are expensive and quickly become stale. Fabraix provides continuous, automated red-teaming that scales with compute, not human effort. By surfacing vulnerabilities before attackers do, it helps organizations build safer, more trustworthy AI systems. The company also publishes research like the Adversarial Cost to Exploit (ACE) benchmark, which measures the token expenditure needed to breach an LLM-backed agent, translating AI security into economic terms.

Launch signal

Fabraix is a Y Combinator-backed company (W25 batch). It has publicly demonstrated vulnerabilities in Fortune 500 agents, published original research (ACE benchmark), and released an open-source Playground on GitHub. The website lists SOC 2 Type II (in progress) and GDPR DPA readiness, indicating enterprise-grade security posture.

Brand and naming

The name 'Fabraix' is a portmanteau of 'fabric' and 'braix' (suggesting 'break'), evoking the idea of breaking through the fabric of AI security. It is distinctive and memorable, though its pronunciation may not be immediately intuitive. The domain fabraix.com is short and brandable. The tagline 'The world's frontier hacker for AI agents' clearly positions the company as a cutting-edge offensive security provider for AI, and the website reinforces this with specific benchmarks and case studies, making the value proposition highly clear.