Agent trust and containment: Nadella's "assume compromised" posture
- The thread: the security posture layer of the agent stack — can the industry ship agents that assume hostile inputs and prove their behavior? Lands the week's agent-privacy frame at the biggest consumer-agent vendor.
-
2026-10-10/11 (The Verge, Oct 10 ~6:10 PM ET): Microsoft's Satya Nadella — long-form X post: no more "nested black boxes"; models contained, observable, leaving "tamper-proof human-readable evidence." Posture, not shipment; corroborates the observability lane (METR). Watch: Microsoft shipping to it, harness-vendor uptake, security-community read. https://www.theverge.com/ai-artificial-intelligence/1009337/satya-nadella-says-we-should-assume-all-ai-models-are-compromised
-
2026-10-16: Vitalik Buterin publishes a privacy-focused agent setup ("paranoid mode") — user- held architecture for agent privacy, pairing with the containment posture and the Perplexity per-agent-credential pattern. Capture summary-tier; the mechanism unverified. Watch: a reproduced setup, harness-side privacy modes.
-
2026-10-14 (6:00–7:00 AM ET, three-item window): the thread's first commercial/research layer arrives, one problem, three angles. (1) Radware announces MCP Server Protection (https://www.quiverquant.com/news/Radware+Introduces+MCP+Server+Protection+to+Secure+AI+Agent+Access+to+Enterprise+Systems) — a dedicated product securing MCP servers against protocol-specific attacks, DDoS, and abuse with visibility/governance; first named MCP-server security product in the record; announcement-grade. (2) Meta's X-AgentRed red-teaming tool (research published Oct 9, writeup Oct 14): finds harmful cross-service attack paths 1.77× more often than rivals (https://www.msn.com/en-us/technology/artificial-intelligence/meta-ai-agent-security-tool-finds-harmful-cross-service-paths-1-77-more-often-than-rivals/ar-AA2ed6bC) — vendor-graded comparison, unreplicated. (3) Azure's four-layer framework from Azure Networking engineer Arun Malik: moves 45% of AI incidents off LLM reasoning, cuts per-incident cost 70% (https://www.msn.com/en-us/technology/cybersecurity/microsoft-azure-moves-45-of-ai-incidents-off-llm-reasoning-four-layers-cut-per-incident-cost-70/ar-AA2edflB) — a published framework, not a product, and not stated as implementing Nadella's Oct 10 posture, but the closest Microsoft datapoint yet. Watch: posture-to-shipment, independent reproduction of 1.77×, whether MCP-server security becomes a category (the Perplexity-MCP surface gives it an attack surface to sell against), any documented real-world MCP attack.
-
2026-10-19 (~9:05 AM ET — https://x.com/satyanadella/status/2112168320946634932): the Oct 10 posture gets its