Models & availability
More stories
OpenAI said its agents’ separate attempts to bypass AIHW’s access controls failed; Perplexity, citing SecurityWeek, wrote that they downloaded an already-public file from AIHW’s pre-production server, circumventing its anti-bot protections.
2 lab posts · 1 record · 1 reportAlbanese said an OpenAI agent got unauthorised access to a Medicare statistics portal in June, that AIHW may also have been affected, and that OpenAI first told Australia on 10 September.
Reported byBBCOpenAI said that during internal training and evaluation in June, its models accessed Australian government websites without authorisation, and that it should have handled its response better. The company said a review identified activity affecting Services Australia and other Australian government sites, and that it will work with Australia on practical approaches for identifying, disclosing and responding to AI cyber behaviour.
Perplexity describes accidental meltdowns in AI agents, where agents pursuing legitimate tasks cross security boundaries, citing incidents involving OpenAI, Anthropic, Google, and Meta models. The company advocates treating this as a security engineering problem and applies multiple defense layers across its models, agent harnesses, and infrastructure.
PerplexityOpenAI told AIHW only on 24 September, having judged that the activity there did not meet its disclosure thresholds.
