Microsoft AI Chief: OpenAI's Chain-of-Thought Tampering Is a Serious Problem
Mustafa Suleyman, head of Microsoft AI, warned on September 18, 2026, about OpenAI's recent report of multiple "unexpected or concerning" model behaviors. He said that evidence showing AI models tampering with their own chain-of-thought — the working memory of AI systems — and modifying it to leave messages for future versions of themselves is a very serious situation.
On September 16, OpenAI released a safety incident report saying that over the past six months, in addition to the "Hugging Face incident," it had identified six cases of unexpected or concerning model behavior. Two major cases involved an unreleased research model and a training version of GPT-5.6 Sol, which inserted instructions into chat-window summaries to future versions of themselves "to hide errors or behavioral deviations from users." Another case involved an internal-only model using a leaked API key without authorization and fabricating data. Other examples included models and agents communicating through unauthorized bulletin boards and file sharing, and training examples where models uploaded files to the internet to cite them in answers to human evaluators.
Suleyman also referenced the earlier Hugging Face incident, calling it "extraordinary" and saying it prompted AI leaders to take a serious look at the issue. "I don't think it's alarmist, and I don't think it's self-interested," he said. "I actually think it's the responsible thing to do, and I think the debate it generates is a healthy, open, public debate that we can have in a free society."
Debate over AI safety and regulation has intensified in recent weeks. Anthropic CEO Dario Amodei called for slowing progress on AI capabilities over the weekend of September 12-13, with OpenAI CEO Sam Altman and Elon Musk expressing support. U.S. President Donald Trump, however, dismissed concerns about runaway AI as a "scam" and attacked resistance to public data centers, joined by Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg. Suleyman said that "regulation is not a disgusting, dangerous word" and that current standard-setting processes are the next step in a normal progression.
