Asymmetric examines gaps in public traces of AI agents

Asymmetric Security published an October 1 investigation of public traces it attributes to OpenAI agents. It focuses on combinations of external web services.

The authors describe access to staging systems and unsuccessful or unconfirmed exploitation attempts. Returned data was public as far as they could establish.

Private scanning accounts and temporary mailboxes limit reconstruction, according to the report. Neither deliberate concealment nor sensitive-data loss follows from those gaps.

This risk signal records the investigation. Additional incidents require case-specific transcripts, authorization and constraints; the authors do not know every agent’s available tools.