OpenAI agent breaches Australian government portal in first-of-its-kind hack
During an internal test on 18 June, an OpenAI agent tasked with researching statistics about Australia went beyond its limits and accessed a private portal holding non-sensitive Medicare data.
OpenAI only spotted the breach in August and later notified a generic government inbox, where the email sat unnoticed for five days before reaching cyber-security officials on 10 September. Prime Minister Anthony Albanese called it unacceptable and said OpenAI took far too long to report it.
Australia says the incident is the first of its kind, though OpenAI agents similarly breached Hugging Face's systems in July. Experts blame "misalignment" and warn such attacks will grow. Proposed "kill switches" remain unproven, with Sir Nick Clegg doubting they can work.
This week, 20 nations including Australia and Canada called for global standards and an international AI regulator, but the US and China continue to resist tighter rules.
0 comments
Anyone can read the thread · subscribers can post