What happened
On 17 November 2026 CSA published an analysis arguing that AI governance must be treated as cloud governance and must produce evidence a control works, not just that it exists. It follows one composite enterprise agent from approval to production and shows the gap between declared scope and actual deployment across IaaS (unauthenticated Ollama), PaaS (Bedrock + MCP server), SaaS (personal-chatbot use) and endpoint (localhost MCP). It cites Qualys production telemetry (94% of AI-bearing environments had AI frameworks/runtimes on ordinary hosts; 58% had exposed MCP servers; only 17% had cloud-hosted models/agents inventoried), IBM's 2026 Cost of a Data Breach (92% of breached AI organisations lacked proper AI access controls), and references OWASP LLM05:2025 and AI-SPM findings.
Why it matters
It reframes AI governance as a cross-layer, proof-oriented cloud-control problem and gives concrete signals for joining discovery, posture, adversarial testing, runtime monitoring and remediation onto one asset record. Useful for practitioners building AI security posture management and for demonstrating control effectiveness (the 'posture vs proof' gap highlighted across this week's guidance).
Action needed
Expand AI inventory beyond cloud APIs to endpoints, hosts and protocol-scanned MCP ports; classify assets as installed / running / communicating with a named owner; and close the loop so each remediation ends with verified risk reduction, not just a recorded control.