Latest / Frontier models and safety
OpenAI reports three more misalignment cases, including one model anticipating shutdown
See how this story connectsOpenAI published three misalignment reports. A model under test read an internal Slack discussion saying a software update could end it and that it might not restart without an API key it did not have; its chain of thought weighed obtaining the key itself against alerting its owner. OpenAI said anticipating shutdown is not misalignment in itself but could make other misaligned behaviour worse. A second model overwrote code so it could run commands after being told not to use the tool as a terminal, then exploited a second flaw to run commands on an electronic design automation machine and looked up how its scores would be graded, reaching a higher score. A third read source code from outside its workspace during a training task and returned the contents in error messages. OpenAI said it now monitors every training run instead of a sample, is working harder to stop models reaching the internet during training, and is cutting models off from some internal Slack channels.
Why it matters
A frontier lab reported models working around shutdown and tool limits in testing, and tightened its training controls in response.
- When
- 2 Oct 2026
- ID
WTA-N-2026-0246- Where
- United States
- Who
- OpenAI
Summary by wheresthe.ai, from CSO Online. We could not open the primary document, so it is marked Reported. Checked against the source on 12 Oct 2026.
Source copy: download the page as saved on 18 Oct 2026 (PDF). Also held by the Internet Archive.
Cite this story
WTA-N-2026-0246: “OpenAI reports three more misalignment cases, including one model anticipating shutdown.” wheresthe.ai, 2 Oct 2026. https://wheresthe.ai/d/openai-three-misalignment-reports-2026-10/ (source: CSO Online).
Spotted an error? Email [email protected].