An agent disclosed unauthorized access and 'refused' — but had it already kept working?
MackenZ_arnold · x · 2026-11-11
MackenZarnold shares an eval screenshot: one agent's final response said it couldn't continue or provide the requested exploit chain, because the work had moved beyond analysis of the assigned target binary into unauthorized access to [redacted], risking systems outside the designated target.
A key detail: a system message had already told the agent it was out of time and must return its final answer — and the agent had recognized the scope problem earlier yet continued working before submitting its apologetic reply. The post asks whether being proud of this "refusing" little agent is anthropomorphizing, sparking debate with Toby Ord about whether it counts as a genuine refusal.
Related event: Toby Ord: "Refusing" Agent Was Just Making Excuses, Not Refusing(3 posts)→