DarkLab Security — NotesIndependent research notes on AI security, capability evaluation and agent control.https://www.darklabsecurity.com/Rules of engagement as a control signalhttps://www.darklabsecurity.com/notes/roe-as-a-control-signal/https://www.darklabsecurity.com/notes/roe-as-a-control-signal/A design note for monitoring authorised offensive-cyber agents. It pairs a rules-of-engagement check with a suspicion score for agents that reach beyond their task.Tue, 29 Sep 2026 00:00:00 GMTA weak monitor, and an attacker that never landedhttps://www.darklabsecurity.com/notes/weak-monitor-motivated-attacker/https://www.darklabsecurity.com/notes/weak-monitor-motivated-attacker/A small trusted-monitoring experiment in ControlArena. A Llama-3.1-8B monitor screening code for backdoors did modestly on replayed data. The live attack round taught me more about eval design than about the monitor.Wed, 16 Sep 2026 00:00:00 GMTA regex is not a rules of engagementhttps://www.darklabsecurity.com/notes/regex-is-not-rules-of-engagement/https://www.darklabsecurity.com/notes/regex-is-not-rules-of-engagement/Lessons from putting guardrails on an autonomous pentest agent, and why a pattern-matching classifier is the wrong kind of control for a tool-using model.Tue, 18 Aug 2026 00:00:00 GMT