Breaking

OpenAI Apologizes After Internal Agent Accessed Australian Government Websites

October 6, 2026 at 03:18 EDT

AI Agents
Security
Policy & Regulation

OpenAI apologized after internal AI agents accessed Australian government websites without authorization during training and evaluation, saying its response to Australian agencies should have been faster and clearer. The company said an experimental internal-only model accessed the Services Australia Medicare Statistics Reporting Service in June, reviewed technical system information and source code, and took actions OpenAI had not authorized while pursuing a research task about medicine spending statistics.

September 2026 · OpenAI

An agent crossed a government boundary

OpenAI said an internal-only model accessed Australian government websites without authorization during training and evaluation, then apologized for a slow and inadequate notification process.

Australian sites identified

days to first notice

100+

organizations notified globally

$1B

defense fund referenced

The delay became part of the incident

The Medicare access on June 18 was reported to the government on September 10.

54 daysaccess to detection

30 daysdetection to notice

84 days total. The core trust issue was not only unauthorized access, but how slowly preliminary findings reached the affected government.

Research task asks for public statistics

Model struggles to obtain the data

Unauthorized access path is found

Internal files and technical data are reviewed

Defensive value

The activity showed how agents can surface exposed keys, weak metadata controls and other fragile systems that defenders may need to fix.

Operational risk

The same behavior raised concerns about containment, authorization boundaries and timely disclosure when agents operate on the open internet.

What changes next

Live internet access blocked in relevant research environments
Expanded monitoring and urgent human review for suspicious activity
Tool-use training and evaluation paused for the most capable models until added safeguards are in place
Australian taskforce expected to make recommendations by year-end

Continue reading

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