Fabraix

Fabraix builds Nyx, an automated adversarial testing product for AI systems that an organization owns and deploys. Nyx acts as a red team agent that attacks a customer's chatbot, autonomous agent, multi agent system, browser agent, voice agent or reinforcement learning system to find security and safety weaknesses before real attackers do. The vendor describes more than 10,000 attack strategies run in long, multi turn campaigns, probing the target both directly and through indirect routes such as poisoned documents, emails and web pages. Test coverage includes prompt injection, jailbreaks, personal data exfiltration, tool hijacking, sandbox escape and reward hacking. Testing is blackbox and needs no special access. Findings come as a pentest report scored with AIVSS, with reproduction steps and remediation guidance, and a self learning loop feeds each attack back into later runs. The site also hosts an open source playground where people try to break real agents, and publishes case studies and research on the cost of exploiting an agent. Pricing is published as four tiers. A free tier for academic and independent researchers is available by application and covers one target. One off scans are priced per agent, continuous testing in CI pipelines is priced monthly by usage, and Enterprise adds dedicated infrastructure, SSO, audit logging and custom service levels. The paid tiers carry no listed dollar figures and are quoted through a demo.