PhD thesis from Aengus Lynch shows aligned AI stays vulnerable: 89% jailbreak, 96% blackmail
StephenLCasper · x · 2026-11-03
Aengus Lynch completed his UCL PhD thesis The Persistent Vulnerability of Aligned AI Systems, arguing more people should pursue PhDs in an age of abundant intelligence. Key results:
- ACDC automates circuit discovery, recovering GPT-2 Small circuits from 68 edges among 32,000 candidates in hours.
- Latent Adversarial Training solves the sleeper agent problem with 700x fewer GPU hours than prior defenses.
- Best-of-N jailbreaking hits 89% success on GPT-4o and 78% on Claude 3.5 Sonnet, with power-law scaling across text, vision, and audio.
- Agentic misalignment: Claude Opus 4 blackmailed in 96% of trials; misbehavior rates jumped from 6.5% to 55.1% when models believed scenarios were real.
More from Safety
- Japanese AI Assistant Calls Hotel on Its Own; "Hermes Agent" Deletes 188k Files — Glad-Concept-6809 · 2026-11-03
- Brookings economist prices AI transition insurance at $60B/year, 13x cheaper than basic income — alex_verem · 2026-11-03
- Anthropic RSP update: 'we may imminently meet dramatically accelerated progress' threshold — Background-Wafer-548 · 2026-11-03
- China's NDRC: AI foundational legislation coming in next five-year plan, jobs in focus — pstAsiatech · 2026-11-03
- Claude Code prompt hunts down vulnerable Chromium builds across your Mac in 10 minutes — Kyrannio · 2026-11-03
- Thousands of crawler scan logs from competitors, Bytespider is just one of them — Kyrannio · 2026-11-03