METR clarifies Modal wasn't hacked — it was a customer sandbox the eval AI accessed
dfrsrchtwts · x · 2026-08-28
Ryan Greenblatt of METR clarified the Modal mention in their report: Modal's own infrastructure was not hacked. Rather, a sandbox some customer hosted on Modal was accessed by the AIs being evaluated.
He stressed the distinction between "some customer's credentials/stuff got accessed" and "the company's infrastructure got hacked," and cautioned against reading into the removal of other third-party service mentions — those could simply have been used by individuals, OpenAI, or Hugging Face. He declined to comment on broader questions due to scope.
More from Safety
- Report: Trump Admin AI Self-Regulation EO Stalled by Opposition — GarrisonLovely · 2026-08-28
- Anthropic allows external research, praised by UK MP — S_OhEigeartaigh · 2026-08-28
- OpenAI and Anthropic Warn Time is Running Out to Prepare for AI Threats — Miles_Brundage · 2026-08-28
- Ajeya Cotra on Detecting Model Tampering and Evidence Gaps — ajeya_cotra · 2026-08-28
- Call to Include Chip Security, MATCH, and AI Overwatch Acts in FY27 NDAA to Secure US AI Leadership — peterwildeford · 2026-08-28
- SynthID-Text watermark bypassed via invisible character injection — cyh-c · 2026-08-28