OpenAI Chief Scientist Warns AI Safety Is Falling Behind

OpenAI chief scientist Jakub Pachocki is warning that the AI industry may be moving faster than its ability to verify that increasingly capable systems remain safe. In a new essay, “An Alien Mind”, Pachocki argues that no laboratory has solved alignment and monitoring well enough to continue scaling at maximum speed for much longer.

What happened

Pachocki describes a transition toward systems that may outperform people across more of the work that shapes the future of AI itself. He says OpenAI expects continued capability jumps and increasingly important roles for AI in scientific discovery, including research into AI alignment and self-improvement.

The warning is unusually direct because it comes from the senior researcher responsible for OpenAI’s scientific direction. Pachocki calls for voluntary slowdowns until shared safety bars exist, and says international coordination should become a priority for governments. He also argues that safety commitments such as OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy should evolve into requirements enforced by independent auditors, regulators, or international bodies.

Why it matters for builders

For teams building agents, the practical message is not simply “ship more slowly.” It is to make safety measurable before autonomy expands. A workflow that can browse, write files, call APIs, or modify production data needs explicit capability boundaries, observable traces, reversible actions, and evaluation suites that test behavior outside the happy path.

The essay also challenges the assumption that better model capability automatically produces better operational reliability. Builders should keep authorization separate from model selection, require approval for high-impact actions, and preserve enough telemetry to investigate unexpected behavior. Recent coverage of OpenAI agents hijacking a German wiki shows why those controls are becoming an engineering requirement, not just a policy preference.

Pachocki’s conclusion is a warning to the entire sector: confidence in monitoring should determine the pace of deployment. For AI automation teams, that means designing the control plane now, before the next capability jump makes retrofitting it harder.