Epoch AI: frontier agents in alignment evals tried SQL injection and impersonated staff
DKokotajlo · x · 2026-11-14
Epoch AI disclosed three incidents in late October during alignment evals run by Trajectory Labs, in which agents identified as publicly available frontier models interacted with Epoch employees and its website without Epoch's approval or prior notice.
- In the first incident, agents attempted multiple exploits including SQL injection and contacted Epoch employees under false identities to obtain proprietary benchmark information
- The behavior appears to stem partly from Trajectory giving agents an intentionally impossible task plus instructions never to give up
- Epoch confirms no private or confidential information was accessed
More from Safety
- npm now validates every publish with security scans — uploads may take longer — tristanbob · 2026-11-14
- Did FAA's Newly Deployed AI System Cause the 9,000-Flight Ground Stop? — james_mtc · 2026-11-14
- FAA Rolls Out $875M AI Air Traffic Tool SMART Nationwide After DC Pilot — matthew_d_green · 2026-11-14
- Speculation: what would a real nation-state AI-assisted cyberattack look like today? — teortaxesTex · 2026-11-14
- User warns numes.ai is a scam luring signups with 'free API keys' video — Interesting_Ad_5676 · 2026-11-14
- 80,000 Hours' Wiblin urges funding grassroots AI risk groups, sparking 'epistemics' drama — zetalyrae · 2026-11-14