Observed incident

Symantec honeypot captures an attempted agent-driven LLM hijack

Reported

Published here

What happened

Symantec captured a request to an apparent exposed AI endpoint carrying an offensive agent’s tool definitions and instructions for scanning, credential theft and mining.

Parties and reporting

AI tools and systems: OpenAI-compatible model endpoint, Offensive agent tool catalogue, XMRig payload

Outcome

The honeypot executed nothing and returned a fallback response, ending the session. It also captured a separate mining-script request; successful downstream compromise was not established.

Operating lesson

Authenticate AI endpoints, limit agent egress and privileges, and monitor unusual inference usage and spending.

Evidence and limits

Observed attempts captured in summer 2026, published October 21. The 100,000-step limit described intended capability, not achieved execution. Supplied training-target history reported no vulnerabilities; telemetry totals are not victim counts. Underlying model and actor identity are unknown.

Sources