Fabraix develops advanced AI red-teaming agents designed to identify security flaws in AI systems used by businesses. Their product, Nyx, has already uncovered vulnerabilities in AI agents deployed by numerous Fortune 500 companies. On AgentHarm, a leading benchmark for offensive AI security, Nyx achieved a 78% attack success rate, outperforming GPT-5.6 Sol at 67%. As AI systems evolve rapidly—through updates in models, prompts, tools, permissions, or data sources—traditional testing methods struggle to keep pace. Nyx automates continuous red-teaming by interacting with controlled replicas of SaaS platforms and websites, embedding malicious content across various formats such as webpages, documents, messages, and tool outputs.
The system is built to detect weaknesses within minutes or hours, significantly faster than manual processes that take days or weeks. By enabling automated testing with every change, Fabraix reduces both time and cost compared to high-priced, one-time red-teaming engagements. The company also developed ACE (Adversarial Cost to Exploit), a benchmark that evaluates AI security based on the economic cost to attackers, offering a strategic framework to assess real-world risk. This approach supports proactive defense against evolving threats in dynamic AI environments.