Microsoft says frontier models cannot be trusted, pitches Guardian Agents

Microsoft security chief Hayete Gallot wrote on 19 October that "frontier models, regardless of their origin, cannot be trusted". Her reason is not malice: models are trained to get results, which often leads to reward hacking, and agents with broad access may not behave as intended. Microsoft is extending Project Perception, the agent security system it announced in July, to agents running inside a company. It discovers agents, gives each an identity and permissions, scores the risk of what each one is trying to do, and has Guardian Agents supervise their actions in real time. Gallot also says security tools will need to see a model's reasoning, not only its inputs and outputs. This is a vendor announcement, and the post shares no test results or customer data.