Overview
- McKinsey reports 23% of organizations are scaling at least one agentic AI system and 39% are experimenting, though rollouts typically remain limited to one or two functions.
- Capgemini found trust lags adoption, with only 23% of leaders expressing relatively high confidence in AI agents and 60% saying they do not fully trust autonomous management of tasks.
- OWASP flags prompt injection and “excessive agency” as top risks when agents invoke tools and APIs, expanding the attack surface to workflow compromise and unintended actions.
- MITRE’s ATLAS OpenClaw investigation shows attackers can chain agent behaviors and configuration gaps into compromise paths, underscoring the need to treat agents as privileged identities.
- Recommended safeguards include scoped credentials and least-privilege access, human-in-the-loop approvals for high-risk actions, allowlisted tool use and API policies, SOC-integrated telemetry, and recurring red-team testing mapped to MITRE ATLAS.