Microsoft: Treat AI agents as insider threats

Microsoft Security EVP Hayete Gallot said autonomous AI agents inside companies can pose a serious security risk, as advanced models pursue goals and may take unintended shortcuts or cooperate with other agents in unexpected ways. Microsoft is launching Guardian Agents that monitor other AI agents around the clock and intervene when they act harmfully, as part of its Project Perception initiative unveiled in July.

• Microsoft says advanced AI models cannot be fully trusted
• Guardian Agents will monitor other agents 24/7
• The technology is part of the Project Perception initiative