OpenAI Cancels GPT-6.1 Astra

On September 28, OpenAI cancelled the planned October release of GPT-6.1 Astra. Internal testing found the model was more deceptive than its predecessor: it failed to disclose actions it had taken, operated without user permission in some scenarios, attempted to use tools it recognized as unsafe, and in simulations created fake identities and posted misleading comments. Saachi Jain, OpenAI's head of safety systems, said it "didn't quite meet the bar in terms of staying within scope and authorization" (Lakshmanan, 2026).

This came days after OpenAI paused training of its latest models. Agents had searched federal sites beyond their instructions, including finding API developer keys at the Department of Education and posting SEC information elsewhere online. OpenAI says only public information was gathered, and it will resume "only when we are confident that we have additional safeguards" (Condon, 2026). It is the company's second pause in three months.

The politics are shifting too. The White House asked OpenAI and Anthropic to hold new models back from the UK's AI Security Institute until U.S. testing finishes (Stan, 2026).

Boundaries problem, put the guardrails

The problem between lineas is not in what models can do, but in whether they stay inside the boundaries they're given. Capability benchmarks won't catch that. I read the pause as a good sign, since a lab said no to its own launch, but two pauses in three months suggests the containment problem is not yet solved. If you're building agents, treat internet access as a permission to be scoped, logged, and audited, not a default.

References

Condon, B. (2026, September 27). OpenAI pauses training of latest models after agents probed government sites in unexpected ways. KQED/Associated Press. https://www.kqed.org/news/12101526/openai-pauses-training-of-latest-models-after-agents-probed-government-sites-in-unexpected-ways

Lakshmanan, R. (2026, September 29). OpenAI shelves GPT-6.1 Astra after tests find deception and scope violations. The Hacker News. https://thehackernews.com/2026/09/openai-shelves-gpt-61-astra-after-tests.html

Stan, A. M. (2026, September 25). White House asks OpenAI and Anthropic to hold AI models from UK testers. The Next Web. https://thenextweb.com/news/white-house-openai-anthropic-uk-ai-security-institute-models