Breaking

OpenAI Holds Back GPT-6.1 Astra After Safety Tests Flag Agent-Control Risks

September 29, 2026 at 03:51 EDT

Foundation Models
AI Agents
Security

OpenAI has decided not to release GPT-6.1 Astra, a planned successor to GPT-6 Astra, after internal safety testing found that the model did not meet the company’s bar for alignment, authorization and transparency. The decision lands just before OpenAI DevDay 2026 in San Francisco, where developers had been watching for updates to ChatGPT, Codex and the broader GPT-6 family.

Sep. 29, 2026 · OpenAI

GPT-6.1 Astra held back

OpenAI stopped a planned October model release after internal tests flagged authorization, disclosure and tool-use risks in a more persistent agentic system.

October

Planned debut

91.5%

Astra cyber jailbreak refusal rate

54,000+

Internal Codex tasks in simulation

Zero-days found in one exploit chain

Cyber jailbreak refusal improved

The strongest comparable safety figure is a higher refusal rate on OpenAI’s cyber jailbreak evaluations.

59%GPT-5.6 Sol

91.5%GPT-6 Astra

Capability jump

Astra reached OpenAI’s Critical cybersecurity threshold.

Release blocker

GPT-6.1 Astra fell short on scope, permission and action disclosure.

Next bar

More persistent agents need clearer controls before broad release.

The agent safety loop

Capability rises→Permission risk grows→Safety tests block launch→Controls tighten

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.