Overview
Platform and organizational policies governing how AI agents access user data, OS permissions, and enterprise systems — spanning consent UX, least-privilege design, and vendor accountability.
Timeline
-
2026-10-23: Tier-one banks (citi, JPMorgan, BofA, others) block third-party agents from transactional account access until know-your-agent standards mature (2026-10-23-citi-cautious-ai-agent-access)
-
2026-10-23: american-express limits external agents to approved API/MCP while restructuring for AI efficiency (2026-10-23-amex-cut-3500-employees-ai)
-
2026-10-03: david-robinson resigned from openai — Atlantic essay criticizes deploy-and-patch safety culture as models grow capable; calls for iterative-deployment-safety reform (openai-david-robinson-resignation)
-
2026-10-03: simon-willison argues hard monthly cloud API spend caps should be default for coding agents — AWS/GCP limits insufficient without opt-out protection (simon-willison-cloud-budget-caps)
-
2026-10-02: apple tightens macOS full-disk-access after muse Messages controversy — first major OS vendor policy response to agent permission abuse (2026-10-02-apple-fda-ai-agents-techcrunch)
-
2026-09-21: muse macOS zero-day demonstrates endpoint blast radius of privileged agents (2026-09-21-meta-muse-zero-day-vulnerability)
Key Players
Analysis
OS vendors are moving from permissive FDA models toward explicit-risk disclosure as always-on agents proliferate. Enterprise buyers should audit agent FDA usage and treat desktop agents as endpoint security surface.