Fabraix

Fabraix builds autonomous red-teaming AI agents that continuously probe customer-facing AI, led by its Nyx agent.

Founded

2026

Product Overview

Fabraix builds red-teaming AI agents that continuously find security vulnerabilities in customer-facing AI, led by its Nyx offensive harness and Arx run-time defence layer.

Nyx is a pure-blackbox autonomous testing harness that probes chat, voice, browser, and coding agents across security, logic, and alignment failure modes. It draws on more than 10,000 jailbreaks and attack strategies and can attack targets directly or indirectly through controlled replicas of SaaS products and websites.

Market Outlook

Fabraix frames AI agent security as a category that static benchmarks and point-in-time audits cannot cover, because non-deterministic agents change behaviour with every model, prompt, or tool update.

The company publishes ACE (Adversarial Cost to Exploit), a benchmark measuring AI security by the token spend an attacker needs to breach an agent, arguing that offensive AI research is the path to safe, trustworthy deployments.

Competitive Advantages

Fabraix reports that Nyx achieved a 78% attack success rate on AgentHarm versus 67% for GPT-5.6 Sol, and positions it as a fully autonomous attacker that profiles, improvises, and chains exploits rather than replaying a static checklist.

Nyx is pure blackbox and needs only an endpoint, URL, or phone number, with no source code, model weights, credentials, or network access. Continuous CI/CD integration and AIVSS-scored reporting reduce the cost and staleness of manual, point-in-time red-team engagements.

Pricing Strategy

Fabraix offers Nyx free to academic and independent researchers by application. Commercial access is priced per agent for a one-off scan and monthly by usage for continuous CI/CD testing.

Enterprise plans add dedicated infrastructure, SSO and audit logging, custom SLAs, and coverage aligned to frameworks such as the EU AI Act, ISO 42001, and MITRE ATLAS.