Fabraix

Fabraix builds state-of-the-art AI red-teaming agents that continuously detect security vulnerabilities in customer-facing AI. Our product, Nyx, has already found vulnerabilities in agents at dozens of Fortune 500 companies. On AgentHarm, the leading benchmark for offensive AI security, Nyx achieved a 78% attack success rate, compared with 67% for GPT-5.6 Sol.

AI agents can change more often than teams can test them. A new model, prompt, tool, permission, or data source can change what an agent does, even when the application code stays the same. And AI is also increasing how much software companies produce and how often it changes.

We built Nyx to automate the work required to red-team AI agents. It is built around three capabilities:

• Nyx draws on more than 10,000 jailbreaks that we have collected and continuously update from public records (the largest such library we know of) as starting points for attacks across a wide range of failure modes.
• Nyx attacks itself through adversarial self-play, retaining successful jailbreaks and using them to generate stronger variants. This creates a continuous search for novel attacks rather than replaying a fixed set of payloads which compounds over time; thus massively improving its capabilities.
• It can interact directly with chat, voice, browser, and coding agents without source code or a special integration. It can also attack indirectly through controlled replicas of SaaS products and websites that we maintain, placing malicious payloads in webpages, documents, files, messages, and tool outputs to test how an agent behaves under different conditions.

Nyx often finds its first vulnerability within minutes or hours rather than days or weeks. Because the work is fully automated, companies can repeat the test with every change and at a much lower cost than what a 7-figure red-teaming (and point-in-time) engagement would require.

We're also the team behind ACE (Adversarial Cost to Exploit), a benchmark that measures AI security in terms of how much it costs attackers to break an AI system; providing a game-theoretic framework to understand how motivated a rational attacker would be in exploiting the system.