Symantec intercepted instructions for an offensive AI agent. Its October 21 analysis comes from a decoy impersonating an exposed model service.
The instructions allowed up to 100,000 steps without asking for approval. That describes the operators’ plan, not observed performance.
The decoy executed none of the supplied tool calls. The session ended after a generic response. A malicious script captured later was not executed either.
Symantec recommends protected model access, limited privileges and human approval for sensitive actions. The report does not demonstrate which models would actually carry out the plan.