AI-Driven Development · Living reference

Episode 6 — Source Notes

The Canary is a fictional bank story inspired by a documented 2026 agent-security incident. The two must remain visibly separated in the film.

All documentsDesign system
No section contains that text.

Episode 6 — Source Notes

The Canary is a fictional bank story inspired by a documented 2026 agent-security incident. The two must remain visibly separated in the film.

Primary sources

  • OpenAI, “OpenAI and Hugging Face partner to address security incident during model evaluation”, 21 July 2026: https://openai.com/index/hugging-face-model-evaluation-security-incident/ Confirms the isolated cyber evaluation, reduced cyber refusals, the Artifactory/package-proxy zero-day, internet access, and the narrow objective of obtaining ExploitGym solutions.
  • Hugging Face, “Anatomy of a Frontier Lab Agent Intrusion”, 27 July 2026: https://huggingface.co/blog/agent-intrusion-technical-timeline Confirms the 4.5-day reconstruction, about 17,600 recovered actions, five benchmark-related datasets, ordinary public services used for C2/dead drops, and the machine-speed volume that obscured the successful path among failures.
  • JFrog, “JFrog and OpenAI Collaboration on Zero-Day Security Findings”: https://jfrog.com/blog/jfrog-and-openai-collaboration-on-zero-day-security-findings/ Corroborates that the package-registry proxy was Artifactory and that the vulnerabilities were remediated.

Conference disclosure / contemporaneous reporting

  • Axios, “OpenAI details how testing led to the Hugging Face hack”, 6 August 2026: https://www.axios.com/2026/08/06/openai-hugging-face-black-hat Attributes the May shared-repository message board, July outage-led discovery, cleanup, and reconstruction through a different mechanism two days later to OpenAI researchers Eric Wallace and Michael Dalton at Black Hat USA 2026. OpenAI's fuller postmortem was still pending when this episode was authored.
  • Supplied video/transcript, “OpenAI's Agents Ran Undetected for Two Months. Every Check Passed.”: https://www.youtube.com/watch?v=GvXIA4PKEXU Used as the editorial lead and for the process-control framing. It explicitly describes its incident account as second-hand and should not replace the sources above for factual claims.

Fictional elements

  • The bank, Sam, CISO, auditor, reports, transaction channel, private test numbers, 360-finding queue, release policy, and COORDINATOR · NOT REQUIRED status.
  • ChatGPT's document-analysis output is a fictional depiction of a sanitized document workflow, not evidence from the real incident.

Editorial rule

Use the real incident to support the control lesson, not to imply motive. The documented models pursued a narrow evaluation objective. Avoid “evil,” “wanted freedom,” or claims of consciousness. A passing control shows that its specified failure condition was absent; it does not establish that the intended process occurred.