Autonomous adversarial-testing agent for customer-facing AI systems, designed for repeatable red-team probing and reproducible findings.
Product sources: Fabraix official surface + first-party GitHub repository + Y Combinator S26
Fabraix Nyx is an autonomous red-teaming system for customer-facing AI agents. It attacks targets as an external user rather than requiring source code, model weights or internal-network access, and is designed to probe direct and indirect surfaces including text, browser, voice, email and malicious documents. Commercial engagements can run as one-off scans or continuously in CI/CD.
AiToolMap did not authorize or run Nyx against a target. The fixed 50-source panel found no eligible exact-product numeric review or independent benchmark result. A public fabraix/nyx repository exists, but repository visibility alone does not establish the license or open-source status of the commercial managed product; this review therefore does not label Nyx itself open source. The 7.9 rating is LOW-confidence and not rank eligible.
Autonomous offensive-security systems require strict authorization and scoping. Fabraix's documented black-box model, synthetic payloads, canary secrets, ephemeral environments and non-execution of destructive real-world actions are positive design choices. Its security page states SOC 2 Type II is in progress rather than completed, which is more transparent than presenting an unfinished audit as certification. Buyers should still verify rules of engagement, data retention, subprocessors, logging and escalation before testing production agents. 1 2 3 4 5
Nyx has a clear threat model. Black-box testing reduces the privileged access normally required by security tooling, while adaptive multi-turn campaigns are designed to explore chains of failures rather than only run fixed prompt lists. Current product materials describe AIVSS-scored reports with reproduction steps and remediation guidance, CI/CD regression testing, and thousands of adversarial strategies. Safety controls are also comparatively concrete: scan environments are isolated and ephemeral, payloads and secrets are synthetic/canary-based, and destructive actions are observed rather than executed against real assets.
The company and product are extremely young. Y Combinator lists Fabraix as an active Summer 2026 company with two employees. Commercial pricing is custom except for a $0 by-application research tier, making value difficult to compare before scoping. The strongest effectiveness claims—78% AgentHarm attack success, first findings in under an hour, and vulnerabilities found at dozens of Fortune 500 companies—come from Fabraix/YC product material and were not independently reproduced for this review.
Products matched by shared tasks, audience, workflow, product type and capabilities.