Fabraix stress-tests AI agents in sandbox

Agent security is shifting from static evals to offensive, runtime testing, and Fabraix is clearly betting that compute-heavy red-teaming beats manual QA for modern agents.

• –Nyx is designed to work without integration, which lowers adoption friction for teams that want to test existing agents as-is
• –The “1,000+ adaptive strategies” pitch matters because agent failures are often emergent and multi-turn, not single-prompt bugs
• –Arx adds a second layer by turning offensive findings into runtime defenses, which is the right shape for a security product
• –The strongest fit is teams shipping tool-using or workflow agents where prompt injection and goal deviation are real production risks
• –The open question is operational trust: buyers will want evidence on false positives, cost, and how well it generalizes across very different agent stacks