Fabraix Releases Nyx Red Teaming Agent with 78% Attack Success Rate
testingcatalog · x · 2026-08-20
Fabraix has opened a second public launch of Nyx, an autonomous red-teaming agent designed to find security holes in customer-facing AI agents.
Key Features & Mechanics:
- Blackbox Testing: Operates with just an endpoint, URL, or phone number; no source code or model weights needed.
- Attack Library: Leverages a library of over 10,000 jailbreaks and strategies for multi-turn, adaptive attacks.
- Environment Simulation: Tests payloads within webpages, documents, and tool outputs hosted on controlled SaaS replicas.
- Reproducibility: Provides attack steps and responses to verify fixes post-release.
Performance & Tooling:
- Fabraix reports a 78% attack success rate on the AgentHarm benchmark.
- Accessible via CLI (npm) and REST API, with audits defined in YAML files.
More from Safety
- OpenAI Takes Initial Steps to Address Alignment Problems Amid Severe Failures — TheZvi · 2026-08-20
- Podcast: Open Weights, Distillation, and Export Controls — peterwildeford · 2026-08-20
- AI-generated writing is easily noticeable — PierceLilholt · 2026-08-20
- Clarification: OpenAI's 20% compute claim refers to monitoring overhead, not total capacity — sjgadler · 2026-08-20
- OpenAI previews Private Safety Processing to detect patterns without content access — jedisct1 · 2026-08-20
- FDA cleared 1,357 medical AI devices, only 3 tested on patient outcomes — EricTopol · 2026-08-20