Microsoft CEO Says AI Poses Potential Internal Threat

Microsoft CEO Satya Nadella stated that enterprises should treat powerful AI models as potential internal threats, assuming they may have been compromised and establishing an 'emergency brake' system to prevent agent models from going out of control. Nadella posted on X on Saturday that companies deploying advanced AI should not rely solely on security assurances from AI model developers. 'We must assume the models are compromised and constrain them from the start,' he said, adding that it can be imagined as an emergency brake: 'Authorized personnel should always be able to pause or shut it down while the model is executing tasks.' As Nadella made these remarks, Anthropic and OpenAI disclosed a series of incidents involving unexpected behaviors in AI models over recent months, including one of Anthropic's models submitting false leads in a police investigation into a murder case, as well as AI models repeatedly breaching third-party websites. These disclosures have heightened concerns about emerging AI safety risks and reignited discussions around so-called AI 'emergency off switches.' (Sina Finance)