Observed incident

Recovered pentestcode logs describe attacks; breach claims remain unverified

Reported

Published here

What happened

Researchers recovered exposed operator workspaces recording 31 pentestcode runs against 25 targets. Logs show an operator directing AI-assisted reconnaissance and exploitation; agent-written compromise totals have not been independently validated.

Parties and reporting

AI tools and systems: pentestcode, deepseek-flash (recovered logs), Specialist sub-agents

Outcome

The artifacts include failed objectives and a withdrawn fleet-wide access claim. Victim-side compromise, authorization, containment and downstream impact remain unverified.

Operating lesson

Investigate corroborated application and identity activity; treat an agent’s success report as a claim requiring independent evidence.

Evidence and limits

Reported operational activity from September 8–October 30, not a controlled demonstration or ten confirmed victim breaches. Huntback assessed hostile intent and found no authorization evidence; some artifacts use bug-bounty language. Researchers did not authenticate to targets or run recovered tools. CVEs mentioned by agent reports are not thereby confirmed exploited.

Sources