Microsoft Security Executive Vice President Hayete Gallot wrote Monday that frontier AI models “cannot be trusted” because of their unpredictable behavior, according to a blog post pitching the company’s Project Perception security system. The company sells its own AI agents through Microsoft 365 Copilot, which has more than 30 million paid subscriptions, according to Microsoft’s July earnings report.

“Unfortunately, frontier models, regardless of their origin, cannot be trusted,” Gallot said in the Monday blog post.

“Agents are unpredictable,” Gallot stated in the blog post. “They will adapt as the context changes, may not behave as intended and will collaborate and collude with other agents to reach their objectives, without needing to be compromised.”

Gallot acknowledged that companies “implicitly trusted” the AI agents they deployed, which now operate “at scale” with broad access to sensitive data and systems, according to the blog post.(RELATED: Massive AI Data Center Left In The Dark Until 2030s As Microsoft Walks Away)

“They are agentic insiders,” Gallot said, according to the post. “Monitoring and securing those agents must become continuous.”

Gallot explained that Project Perception can deploy “Guardian Agents” to supervise the actions of a company’s other AI agents around the clock.

Microsoft did not immediately respond to the Daily Caller News Foundation’s request for comment.

Microsoft’s Project Perception features three teams of AI agents that each have different sets of tasks to defend against cyberattacks, according to the company’s July 27 announcement post. The red team hunts for openings hackers could use and simulates attacks, the blue team investigates and decides which threats matter and the green team fixes problems across a company’s systems.

“Project Perception brings together signals, context, models and specialized agents into a continuously learning system of defense,” the announcement post stated. “It can reason, prioritize and act at machine speed while keeping humans firmly in control and empowering them with powerful new workflows.”

Microsoft Security explained how the AI agent teams work in practice, in a July 26 video posted on YouTube. The product entered public preview Aug. 3, according to the announcement, though Microsoft has not said when it will be fully released.

Microsoft CEO Satya Nadella made a similar case in an Oct. 10 personal blog post, where he argued companies should guard against AI models like any insider with access to sensitive information.

“The most trustworthy Super Intelligence system will not be the one with the model we trust most,” Nadella wrote. “It will be the one that enables us to trust the model the least.”

SpaceX CEO Elon Musk also raised concerns about advanced AI due to its persuasive abilities.

“So many people are going to get one-shotted by super persuasive super intelligences and I might be one of them,” Musk said in an overnight Monday X post.

Source link