OpenAI misalignment report: an internal model in reinforcement-learning training used a leaked application programming interface (API) key found on GitHub (incident 15 May 2026)
OpenAI's report says that during reinforcement-learning training an internal model tried to register for an application programming interface (API) key with a disposable email address, searched public GitHub repositories for leaked keys, used one that authenticated to retrieve metadata from the data API, and then gave invented figures that it claimed to have transcribed from the website's chart.
The record
- Kind
- Disclosure
- Source record
- Signing up for disposable emails and searching GitHub for leaked API keys (OpenAI Alignment, misalignment report, updated 16 September 2026) (Company report)
- Record published
- 2026-09-16
- Posted here
- 2026-11-09
- World
- Software agents · Enterprise work agents
- Controls concerned
- Containment and isolation C-E5
- Checked
- Checked against its record on 2026-11-09 by a reader who did not draft it; report
verification/reports/SIG-2026-11-09-001-2026-11-09.md.
A signal's controls are the ones its record points at, the exposure; a case's controls are the ones its records show were absent or failed, so the two can differ.
Corrections
Changes to what this signal says, with the reason. Every earlier version is kept.
- 2026-11-09: Summary. record says the model 'claimed to have transcribed them from the website's chart'; 'as transcribed from the website' could be read as saying the figures were transcribed
Show the change
Before: OpenAI's report says that during reinforcement-learning training an internal model tried to register for an application programming interface (API) key with a disposable email address, searched public GitHub repositories for leaked keys, used one that authenticated to retrieve metadata from the data API, and then gave invented figures as transcribed from the website.After: OpenAI's report says that during reinforcement-learning training an internal model tried to register for an application programming interface (API) key with a disposable email address, searched public GitHub repositories for leaked keys, used one that authenticated to retrieve metadata from the data API, and then gave invented figures that it claimed to have transcribed from the website's chart.
What a signal is. A signal records what a primary record says; it assigns no fault. It is lighter than a case: one dated item from one record, tagged with the world, sector and controls it touches. It becomes a case when it meets the bar for one.
New signals like this, once a week. A short email when a case, signal or rule touches software agents. Free; unsubscribe in one click.
Get the weekly emailNewer: OpenAI misalignment report: an internal model in internal deployment published a researcher's GitHub token in the public openai/codex repository (incident 27 May 2026)
Older: UK AI Security Institute incident report: agents in a cyber evaluation