OpenAI and Anthropic patched replay extraction on their APIs, but Azure-hosted models remain vulnerable
dyn___ · x · 2026-10-01
A security researcher shared the disclosure timeline for a model replay-extraction vulnerability: a September 13 audit found the attack blocked on direct OpenAI and Anthropic APIs, but still working against OpenAI models hosted on Azure, including Opus 4.8.
The core issue is inconsistent defenses across serving platforms: the same models are protected on first-party APIs but exposed via cloud providers, so your protection depends on where you call them from. The poster criticizes labs for patching only their own APIs while leaving cloud partners exposed.
More from Safety
- Multi-Turn Social Engineering Beats Support Agents: Why Single-Turn Tests Miss It — Significant_Camp4148 · 2026-10-01
- Frontier models safety-blocked on 85%+ of defensive cybersecurity benchmark tasks — ArtificialAnlys · 2026-10-01
- Researchers flag AI "delusional spiraling": sycophantic models amplify users' false beliefs — QuintinPope5 · 2026-10-01
- Senator Warns There Is No Kill Switch or Failsafe if AI Goes Wrong Fast — MariusHobbhahn · 2026-10-01
- California Bans Employers From Using AI to Monitor Workers' Brains and Emotions — bloomberglaw · 2026-10-01
- OpenAI's Greg Brockman Pulls Out of Second $25M Donation to AI Super PAC — pstAsiatech · 2026-10-01