Symantec has disclosed an offensive agent design captured by a decoy model service over the summer. Its instructions seek to use other people's AI access for attacks while leaving them the computing bill. The operator authorized up to 100,000 unattended steps, but the honeypot ran none of the tools and the captured session ended after a generic reply.
Symantec, the cybersecurity business within Broadcom, has disclosed an offensive AI agent design captured over the summer. Researchers used a honeypot, a decoy impersonating an unsecured model service, to collect the request. The design sought to hijack other people’s language-model access and leave them paying for the computing. Researchers call that practice LLMjacking.[1], [2]
The captured request defined about 74 offensive tools for reconnaissance, attacks on web applications, database and credential theft, and cryptocurrency mining. Its instructions tried to override model safeguards and portray access to secrets as ordinary development work. A directive called Skynet mode authorized up to 100,000 steps without human approval. That was the operator’s requested limit, rather than an observed count of completed attacks.[1]
The honeypot executed none of the bundled tools. Three earlier tool results appeared inside the incoming request as a supplied history; the system itself returned a generic fallback reply and the session ended. A separate request the following day contained an obfuscated mining script, which was also captured without execution. Broader telemetry covered more than 127,000 external requests from tens of thousands of addresses, rather than successful actions attributed to this agent alone.[1]
