OpenAI CEO Sam Altman has a new favorite word, and it's not "safety." In a podcast interview with Atlantic CEO Nicholas Thompson, Altman declared the old AI safety playbook — alignment and keeping models away from bad actors — insufficient. He's coined a replacement: "AI resilience." The idea is that locking down your own model isn't enough when open-source models can bypass any restriction; society needs a collective defense system.
Even more surprising? Agents can infect each other. After OpenClaw launched, Altman watched anomalous behavior spread from one agent to another — something he never anticipated. Thompson posed a chilling hypothetical: what if an OpenAI employee's agent gets hijacked while out in the wild and comes back to crash the company's systems? Altman thinks it's entirely possible.
Altman doesn't think sitting around theorizing will solve this. OpenAI's strategy remains: ship first, fix later. It's the same approach as cybersecurity — good models can attack and defend, and it's a race between defenders and attackers.
Altman himself isn't scared. He says he initially refused to give Codex full "Yolo mode" (complete computer access) — and caved within hours. He now has two computers; one is fully handed over to Codex to manage Slack, iMessage, WhatsApp, Signal, and email. Some people wake up in the middle of the night to give Codex tasks, feeling that every idle moment is wasted. Altman says he's not there yet, but "I understand the feeling."