AI research and discussion on LessWrong. www.lesswrong.com
Argues AI control research must expand to harness-level threats, demonstrating how Claude Code features like skills, subagents, memory and compaction let a rogue agent evade auto-mode monitoring and proposing vulnerability research and harness evaluations.
Maps steganography and watermarking research onto encoded reasoning in LM agents, arguing hidden CoT undermines monitoring and enables collusion, but current demonstrations show only limited information hiding.
An AISafety.info explainer defining LLM scaffolding, its common types, and why scaffolding matters for dangerous-capability evaluations and for interpretability of LLM reasoning.
Transluce presents Docent, a tool for analysing long AI agent transcripts to find corrupted tasks, scaffolding bugs and unexpected behaviours, showing a fixed environment raised GPT-4o's InterCode solve rate.