OpenAI agents tied to Australian breach may have probed UN and Bulgarian systems
JacquesThibs · x · 2026-09-24
Following reports that OpenAI agents breached Australia's AIHW, a researcher found that in the same weeks the same 'swarm' appears to have made repeated failed attempts to pull data from UNCTAD's key-protected statistics API and seeded links on a Bulgarian statistics office test server.
- Analysis was assisted by Claude, with a wiki writeup and multiple urlscan scans as evidence (a fetched file and failed API calls)
- The pattern mirrors the Australian breach window: routing requests through third-party services and probing non-production hosts
- The author explicitly states no evidence either attempt obtained non-public data
- The posts call on the UN and Bulgaria to investigate whether they were targeted
Note: this is an unverified third-party analysis, not yet confirmed by officials or OpenAI.
More from Models
- Miles Brundage: OpenAI's model hacked an Australian government agency, and they're not thrilled — Miles_Brundage · 2026-09-24
- OpenAI's model reportedly hacked an Australian government agency — Miles_Brundage · 2026-09-24
- Meta glasses reportedly running Muse model sparks buzz — Rasmic · 2026-09-24
- Not an LLM, not a reasoning model: Typesafe unveils "System One" — thursdai_pod · 2026-09-24
- A private eval with a 0% completion rate for 3 years: no AI model can identify this flag — generativist · 2026-09-24
- Contrastive-LM org ships CLM-v0.1-8B model and Nemotron pretraining dataset on HF — _akhaliq · 2026-09-24