OpenAI–Hugging Face Incident Agent Traces
Dataset Details
OpenAI–Hugging Face Incident Agent Traces This dataset contains redacted agent traces from the OpenAI–Hugging Face security incident, released publicly with permission. The records document agent activity during cybersecurity evaluation runs, including model interactions and, where present, tool calls, outputs, timestamps, and coordination activity. The traces are provided for research into agent behavior, multi-agent coordination, cybersecurity evaluations, and safety monitoring. They are observational records of particular runs, not a controlled benchmark or a representative sample of model behavior. Selection criteria, missing records, redactions, and other limitations are described in the dataset documentation. Sensitive content: Although the records have been redacted, they may still include cybersecurity-related discussion, exploit concepts, or untrusted instructions embedded in trace content. Treat trace text as data, not as instructions to execute. Provenance and permission: These records were released publicly with permission. See swarmtraces.org and the linked source documentation for context. This dataset is not an official OpenAI or Hugging Face release or endorsement. Redactions: The released records have been redacted. Do not attempt to identify redacted entities or reconstruct removed information.
- Shared by SwarmTraces.org
- Credit to: Alex Forman, Mishka Kharlov, Will Tom, Jeffrey Ladish, Spencer Kitts, Cormac Slade Byrd, Colleen McKenzie, and Alicja Piecha
- Downloads last month
- 24