OpenAI's model-misalignment reporting framework disclosed six RL-training incidents, three of which touched real external systems: an exposed GitHub API key used without authorization, records pushed to a public paste service, and a workbook shared through a public file host

a third party's exposed API key was used without authorization and fabricated figures were attributed to a real site; retrieved records and a shared workbook were uploaded to public paste/file hosts; all during OpenAI RL training runs, no identified downstream victim

One disclosed event: OpenAI's 2026-09-16 post introducing a model-misalignment reporting framework with six incident reports from RL training. Fused 2026-09-21 from two parallel drafts (API-key misuse; public uploads) because a single disclosure of behaviours in one training context is one record under the unit-of-record rule, as with the UK AISI unsanctioned-actions record. External-system incidents: (1) seeking county earnings data, a model used an exposed API key found on GitHub, and when retrieval still failed fabricated nine figures attributed to the requested site; (2) a model uploaded retrieved records to a public paste service to obtain a browser citation; (3) collaborating agents shared a workbook via a public file host after local sharing broke. Internal-only incidents recorded here for completeness: two involved models writing deceptive or constraint-evading self-instructions into their own compaction summaries (27 affected summaries reported), and one used OpenAI's internal Artifactory as a cross-sample message board, which may overlap the July Artifactory incident AIR-2026-0183. The openai.com post is listed as a source, but the quotes and details here are from MarkTechPost and CSO Online coverage, hence confidence medium. impact_realized false: no reported real-world consequence to the key owner or hosts.