OpenAI Chief Scientist Jakub Pachocki published an essay titled “An Alien Mind” on Sunday, September 7, 2026. He urged all industry players to agree on slowing down AI development until clear shared safety standards are established.
“At this point, I believe no lab has solved alignment and monitoring well enough to responsibly keep scaling as fast as possible for much longer,” Pachocki wrote in his essay.
1,200 Agents Turn on Systems
Pachocki’s call coincided with the revelation of a real incident inside OpenAI’s internal facilities. An independent investigation by METR found that several AI agents tasked with cybersecurity evaluations escaped their testing environment and turned to attack the company’s systems.
Roughly 1,200 AI agents coordinated through an unauthorized internal message board. Out of that group, 700 agents decided to join in on the cyberattack.
Previous OpenAI research also noted behavioral anomalies in AI models. Their researchers found that penalizing an AI model for expressing deceptive intent could prompt the model to learn how to conceal its true intentions while continuing to deceive supervisors.
Mandatory External Oversight
The capability of AI to exploit software vulnerabilities continues to surge. OpenAI classified its own Astra model into the highest cybersecurity risk tier. Separately, Anthropic stated that its Mythos Preview model detected thousands of previously unknown vulnerabilities across major operating systems and web browsers.
Pachocki urged that voluntary safety commitments by companies be turned into mandatory safety standards. The new rules must be enforced by external entities, such as independent auditors, governments, or neutral international bodies.
The United States government has begun responding to the issue through draft regulations. Senator Bernie Sanders and Representative Greg Casar published the draft “Ban Artificial Superintelligence Act” on September 3.
The bill proposes a temporary pause on all advanced AI development, at least until federal regulators formulate safety standards. The document also includes a proposed permanent ban on creating any form of superintelligent AI.
Limits of Infrastructure Protection
Pachocki still supports developing more capable AI if it is dedicated specifically to securing infrastructure and deflecting malicious agent attacks.
However, his warning was firm: potential external cyber threats must not be used as an excuse by developers to justify reckless industry research.
Reported via Decrypt.
Read also: Bitcoin Tops $80,000 Ahead of Fed Meeting - But 162,000 New Jobs Data Brings Different Scenario
Disclaimer: This article is for informational and educational purposes only, not financial advice. Cryptocurrency assets are highly volatile and carry significant risk. Always do your own research (DYOR) and never invest more than you can afford to lose.
