Microsoft published two security blog posts on October 19, 2026, warning enterprises about autonomous AI agents, which it labels as “agentic insiders”. These agents now have extensive access to business systems and data, often with more privileges than they need, according to Hayete Gallot, Microsoft’s EVP of Security. The company launched Guardian Agents, monitoring tools integrated into Project Perception, capable of restricting permissions or requiring human validation when risks are detected. The number of detectable agent types increased from 25 to 60, including widely used tools like OpenAI’s Codex and GitHub Copilot, a Microsoft subsidiary. CEO Satya Nadella summarized the approach by stating that the most secure systems are those that place the least trust in their AI models.
Source: Read the original article
