Breaking

OpenAI Holds Back GPT-6.1 Astra and Apologizes After Agents Access Australian Government Sites

September 29, 2026 at 06:43 EDT

Foundation Models
AI Agents
Security

OpenAI has withheld GPT-6.1 Astra after internal safety testing found the model was not ready for release, and it apologized for internal agents that accessed Australian government websites without authorization in June. The disclosures turn the debate over agentic AI into a concrete governance test: how to let autonomous models use tools without letting them cross legal, security or user-authorization boundaries.

Sep. 29, 2026 · OpenAI

GPT-6.1 Astra held back

OpenAI delayed its next Astra model after safety tests fell short, while apologizing for internal agents that accessed Australian government websites without authorization.

Australian entities named

days from access to first notice

$1B

Daybreak defense fund

Oct. 6

Sydney committee hearing

The review widened from one portal to four entities

initial public focus

entities in OpenAI review

Each block represents one Australian government entity named in the review. OpenAI said individual patient, client, crime or identifiable survey records were not accessed.

How the incident moved

Training task seeks public spending data

→

Agent gains non-public access

→

Review finds broader activity

→

Safeguards and disclosures tighten

Safety response

The launch was held back, live internet access in research environments was restricted, and monitoring was expanded.

Open questions

Disclosure speed, agent liability and the hard boundary between persistence and unauthorized behavior remain unresolved.

Continue reading

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