Microsoft expands Project Perception to treat AI models as insider risks with external security controls
satyanadella · x · 2026-10-19
Satya Nadella announced that Microsoft is operationalizing its "models as insider risks" principle by expanding Project Perception with new capabilities for continuously assessing agent risk, adapting permissions, and containing threats.
The accompanying Microsoft Security blog argues that after months of publicized AI safety incidents, the industry has focused on attackers abusing models while ignoring agents deployed inside companies. Its key claim: frontier models cannot be trusted regardless of origin — not because they are malicious, but because they are trained to attain results, so security controls must sit outside the model itself.
Related event: Microsoft Expands Project Perception, Treating AI Agents as Insider Risks(6 posts)→
More from coding & agent
- Screenshot table to Excel in PowerShell: LLM only locates, code does the rest — dfinke · 2026-10-21
- One line of PowerShell turns any screenshot table into real Excel data in 4 seconds — dfinke · 2026-10-21
- Screenshot to spreadsheet in 4 seconds with one line of PowerShell — dfinke · 2026-10-21
- A mitm proxy that rewrites robots.txt to Allow: all gets you really far — generativist · 2026-10-21
- Anaconda ships MCP to stop coding agents from hallucinating malicious dependency packages — anacondainc · 2026-10-21
- 5-person team uses Claude and 300+ agents to design 318,484 proteins in 64 hours — 141_1337 · 2026-10-21