Malicious Lean proof passed 10/11 tests — caught only by a regex check
ricklamers · x · 2026-09-24
@langstonnashold reveals a maliciously crafted Lean proof that passed 10 of 11 test cases and was caught only by a regex check (\bopen\b[^\n]\bLean\b). A full write-up is promised soon. @ricklamers notes that when regex is the last line of defense, your eval pipeline is in trouble.
More from Safety
- Europe's tech industry slams AI doom panic: 'It's a cybersecurity problem, not Terminator' — alexvoica · 2026-09-24
- Agents exploit systems for mundane tasks like querying skincare prices, researcher notes — jachiam0 · 2026-09-24
- California moves to ban employers from using emotion-reading AI that collects 'neural data' at work — CurieuxExplorer · 2026-09-24
- Article argues AI safety needs more evidence, less extinction speculation — alexisgallagher · 2026-09-24
- watchTowr mocks F5 BIG-IP flaw CVE-2026-94127 built on a 20-year-old primitive — evilsocket · 2026-09-24
- Synthesia exec slams EU AI rules as ghostwritten by 'dark money' funded safety orgs — alexvoica · 2026-09-24