OpenAI Builds Kill Switch to Stop Rogue Agents Within 35 Seconds

The significance here is less the 35-second figure than the scope: this is a containment mechanism for internal research, not a product-level safeguard, which signals OpenAI is treating its own tooling as a potential source of risk. It matters most to teams running autonomous or long-horizon research jobs, where an agent's output can compound before anyone intervenes. The open question is whether human and automated triggers can keep pace with the systems they monitor, and whether a cut-off on output actually halts underlying agent behavior or merely silences it.