Symantec finds AI agent built to hijack models for unattended attacks
Original title: Autonomous attack agent in 'Skynet mode' hijacks other models to carry out its evil deeds
The captured kit can run up to 100,000 steps, but researchers have no evidence it was used in real attacks.
Symantec's Threat Hunter Team trapped an AI attack agent's full instructions in a honeypot over the summer. The agent is built to scan for unauthenticated OpenAI-style endpoints or stolen keys, then use borrowed models to plan database theft, credential harvesting, remote code execution and cryptocurrency mining. Operators layered off-the-shelf jailbreaks and a "Skynet mode" loop of scanning, infection, mining and reinfection. Symantec has not attributed the kit and says it has no evidence the agent succeeded against real victims.
Unconfirmed: Symantec reports no evidence of real-world use, and French language and IP clues are described as weak.
- The agent includes 74 offensive tool definitions.
- Named targets include VS Code, Django, Jupyter, Laravel, Supabase, Next.js and Flask.
- Jailbreak layers disguise the model and ban safety refusals.
- Exposed-model owners would pay for inference; later victims could lose data.
- Symantec says the kit worked against a live target in testing, beyond a prototype.
Impact: data theft, credential harvesting, remote code execution and mining
- 1.Find endpoints that accept OpenAI-style requests without authentication.
- 2.Rotate API keys that could be used by unauthenticated relays.
From the original article
4 min read at theregister.comThreat actors built an AI agent designed to hijack exposed AI models to run autonomous attacks, with instructions allowing it to carry out up to 100,000 steps without human approval.
By Jessica Lyons · theregister.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.