In an Oct 16 ISC diary, Mark Baggett argues the industry is approaching agent security backwards by wrapping powerful general agents in containers, guardrails and approval prompts, which he doubts will hold, especially when the underlying model is swapped out. His alternative is to use coding agents to build narrow agents that only get the tools and paths their job needs. He cites Check Point's Oct 15 'Dangerous Deviation Detection' beta, which checks a proposed tool call against the task before it runs and flags actions that are both dangerous (data leakage, privilege escalation, destroying resources) and unwarranted. His example: an agent asked to investigate a login that goes on to disable accounts. This is opinion and also promotes a SANS course, but the scope-creep failure mode it describes needs no attacker at all.