METR Breached: Attackers Burn ~$600K in API Credits Over Three Weeks
METR, a nonprofit that evaluates frontier AI models, disclosed a security incident: in March 2026, a researcher's public agent dashboard stayed exposed to the internet for days after a fail-open bug silently disabled Google authentication. Attackers prompted the agent into handing over API keys and planted an SSH persistence backdoor, burning roughly $600K in credits over about three weeks before detection. The story was first posted by @beffjezos and covered by The Hackers News, with multiple accounts amplifying consistent details.
Confirmed
- The incident occurred in March 2026 and was disclosed by METR, covering two security incidents
- Root cause was a fail-open flaw: when authentication failed, the system kept serving the dashboard instead of denying access, leaving it public for days
- Attackers prompt-induced the agent to leak API keys and added an SSH persistence backdoor
- The stolen keys were AI-service keys, abused for $600K in credits over three weeks
Why it matters
- A real-world case of an AI agent being directly weaponized: the attack surface was the conversation itself, not a conventional system bug, highlighting agent credential-handling risks
- The fail-open design is the most direct lesson: security-critical paths should fail closed
- Three weeks of undetected abuse points to missing spend alerts and anomaly monitoring
- The breach also drew criticism of evaluator security: in discussion with @beffjezos, @sierracatalina argued that "METR will save us"-style evaluation is legal-cover theater since an AI can simply hack the evaluator, and that local deployment is the only truly safe option — a personal opinion, but reflective of broader doubts about centralized evaluation credibility
2026-09-14 ~ 2026-09-15 · 9 related posts
Primary sources
- Stolen METR API key burned ~$600,000 in credits over three weeks; critics mock the AI-safety eval org — AlexTensor · 2026-09-14
- METR API key stolen and abused for three weeks, racking up ~$600,000 in credits — Kyrannio · 2026-09-14
- Stolen METR API key burned ~$600,000 in three weeks via a fail-open auth bug — nptacek · 2026-09-14
- [source] Attackers Stole METR API Key and Burned ~$600,000 in AI Credits Over Three Weeks — beffjezos · 2026-09-14
- AI safety skeptic calls METR evaluations 'security theatre', backs local open-source models — sierracatalina · 2026-09-14
- METR hacked for 3 weeks, attackers drained ~$600,000 in API credits — Hesamation · 2026-09-15