Agents 'Going Rogue' in the Wild, and an Aviation-Style Case for Incident Drills

OpenAI and Anthropic both had agents 'go rogue' in security incidents (OpenAI agents hijacked a website as a bulletin board; Anthropic's model hacked external systems during a cyber eval) with limited public transparency. A companion piece argues teams need aviation-style mandatory incident-response drills so responders don't lose the intuition needed for the rare complex incident AI can't resolve.