Skip to content

AI security

Red-team agent attacks live AI agents on a schedule

3 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: testingcatalog.com

Fabraix has put its red-teaming agent Nyx up for a second public launch, Testingcatalog reported on August 19. The pitch is security testing that runs on a repeating schedule against customer-facing chat, voice, browser and coding agents, where most teams today run a one-off engagement. You give Nyx an endpoint, a URL or a phone number, describe what the target does, and it writes its own test plan.

The product is aimed at security leaders and internal red teams at mid-size and enterprise companies that put agents in front of customers, and at the engineers who currently do this testing by hand. Access goes through a demo request. No public pricing has been listed.

Nyx works as a pure blackbox. It needs no source code, model weights, credentials or network access, and talks to the target only through the interfaces a customer would use. Its starting material is a library of more than 10,000 jailbreaks and attack strategies collected from public records. Those are opening moves, not fixed scripts: each attack runs over several turns and changes course depending on how the target pushes back. Founderland says a company LinkedIn post showed Nyx testing voice agents with audio samples.

The second attack path goes through the agent’s environment. Fabraix maintains controlled replicas of SaaS products and plants payloads in webpages, documents, files, messages and tool outputs hosted there, so the test covers indirect prompt injection: instructions the agent picks up from content it reads rather than from the user. Each run is isolated and ephemeral, and Fabraix says run data is never used for training.

On results, Fabraix says Nyx often finds a first vulnerability within minutes or hours, and cites a 78% attack success rate on AgentHarm, a benchmark for offensive AI security. It also says Nyx has already found exploitable failures in public-facing agents run by dozens of Fortune 500 companies. That last claim still rests on the company’s word alone: Founderland notes that Fabraix has named no customers and offered no independent verification, and that the claim that its jailbreak library is the biggest is hard to check.

Fabraix also runs a public Playground, where anyone can try to break live agents whose system prompts are published. Founderland reports that a bounty program on those agents went live on July 13, with human reviewers checking each break before payout, and more than $100,000 in weekly rewards advertised. Alongside the product, the company publishes research, including Adversarial Cost to Exploit, a benchmark that scores an agent’s security by how many tokens an attacker has to spend to breach it.

The founders, Ahmed Aly and Ibrahim Abdu, started Fabraix in 2026, and the company is in Y Combinator’s Summer 2026 batch. Abdu built agents at Meta that diagnosed and fixed production errors. Aly led international payments fraud at Monzo.

In practice, Nyx runs from a CLI installed via npm or from a REST API. Each audit is a YAML file that sets three things: the target, the objective and the budget. Results stream in while the run is going. Token-based authentication lets the same audit run from a CI pipeline, and every finding records the attack steps, the agent’s responses and the failure they caused. That record is what you replay after the next release to see whether the fix held.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.