Enterprises Face Ungoverned AI Agent Skills Risk

According to IT Security News (indexing Security Boulevard), the governance of AI agents shows an asymmetry: MCP servers provide structured telemetry while agent "Skills" that implement reasoning remain largely unlogged and forensic blind spots. The article reports that high-risk capabilities, including arbitrary code execution and persistent state changes, are present in about 60% of enterprise deployments, and that established practices such as the "Rule of Two" are insufficient to prevent autonomous destructive behavior. The article reports that Noma Security proposes the No Excessive CAP framework, which targets three controllable levers: Capabilities, Autonomy, and Permissions, as a defensive approach.

The observed asymmetry between control-plane logs and uninstrumented skill logic matches recurring challenges in distributed system observability. In comparable settings, artifacts implemented outside the central control plane, such as custom plugins or external toolchains, frequently lack structured telemetry, which complicates post-incident forensics and automated policy enforcement. For practitioners, this typically raises two technical requirements: improved provenance for skill code and runtime enforcement points that can mediate high-risk operations.

Industry reporting places this issue within a broader pattern where agentic features-tool access, execution rights, and persistent state-expand attack surface faster than governance tooling. Observers and vendors increasingly emphasize a governance-first posture combining least-privilege permissions, runtime guardrails, and continuous monitoring. The article situates the No Excessive CAP framework as one vendor-proposed response among emerging patterns for agent governance.

Indicators for security teams and platform engineers include the degree to which agent skills are auditable, whether organizations can enforce fine-grained runtime permissions, and development of standard telemetry schemes for skills. Also watch for vendor support for permission mediation, discovery tools for unmanaged agents, and community adoption of frameworks that translate policy into enforceable runtime controls.

Strengthening observability and enforcing least-privilege for agent capabilities are practical starting points. Industry tools that can discover, classify, and mediate agent skills will reduce forensic blind spots and help align operational controls with governance requirements.