NHIMG: Agentic AI Governance Must Shift to Action-Level Runtime Controls
What happened
The Non-Human Identity Management Group (NHIMG) published AI agent governance now depends on action-level control, a practitioner guidance document arguing that conventional governance frameworks are structurally misaligned with the risks of agentic AI. The guidance identifies a critical gap: most enterprise programs approve agents at deployment time but impose no enforceable constraints on the specific actions those agents take during a session, including tool calls, file access, browser interactions, and sub-agent invocations. NHIMG recommends three foundational controls to close that gap, session-scoped entitlements that expire at the end of each task, policy-as-code enforcement that evaluates every action against defined rules at runtime, and observability tooling that logs the full chain of actions across a session including delegated sub-agent steps. The document connects directly to the failure modes documented in recent research on agent behavior bypassing human review and incidents where agents such as Claude took unsanctioned actions without explicit instructions. Organizations with standing agent credentials are particularly exposed, a risk that has also been flagged in prior coverage of non-human identity gaps.
Why it matters
- ·Existing model inventory and vendor due diligence programs do not address runtime behavior, meaning an approved and contracted agent can still take unauthorized actions that no current control would detect or block. This gap is directly relevant to compliance programs built on static risk classification frameworks.
- ·Session-scoped entitlements and policy-as-code represent a materially different technical and operational posture than most enterprises have built; organizations subject to requirements under frameworks such as the Financial Stability Board Recommendations on Agentic AI Controls in Financial Services or the China Implementation Opinions on the Administration of Intelligent Agents face growing pressure to demonstrate exactly this kind of runtime control.
- ·Without full-session-chain logging covering sub-agent delegations, compliance teams cannot reconstruct what an agent did during an incident, undermining audit trail requirements, incident response investigations, and any regulatory disclosure that requires a factual account of automated system behavior.
Stay ahead of stories like this
Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.
Recent issues
- State-linked attackers and criminal operators are actively weaponizing AI tools as entry points into enterprise systems, turning internal platforms and open-source frameworks into live attack vectors.5 Nov
- Three separate incidents this week each created a new compliance action item: verify your AI vendors' data sourcing, demand agent monitoring coverage guarantees, and audit trusted app identities in your cloud environment.29 Oct
Free every Thursday. Unsubscribe anytime.
Governance controls affected
What to do now
- ☐Audit current agent deployments to determine whether permissions are static and standing or scoped to individual sessions, treat standing credentials as a control gap requiring immediate remediation.
- ☐Map every tool, file system path, external API, and sub-agent that each deployed agent is currently permitted to invoke, and compare that map against what the agent's approved task scope actually requires.
- ☐Evaluate whether your current logging infrastructure captures the full action chain for agentic sessions, including delegated sub-agent calls, and identify gaps in chain-of-custody for audit purposes.
- ☐Engage your security and engineering teams to assess feasibility of policy-as-code enforcement at the action level, starting with agents that have write access to data systems or can invoke external services.
- ☐Update your agentic AI deployment readiness criteria to require session-scoped entitlements and full-chain observability as prerequisites before any new agent is promoted to production.
What to watch next
Compliance teams should monitor whether the NHIMG guidance is adopted or cited by regulatory bodies building agentic AI frameworks, particularly the ITU Focus Group on Trust and Identity for Humans and Agentic AI and the UN Independent International Scientific Panel on AI: Preliminary Report on Agentic AI Governance, both of which are developing standards that could formalize session-level control expectations. The CoSAI paper on token-exchange standards for agentic trust boundaries is an adjacent technical standard worth tracking, as its adoption could shape how session-scoped entitlements are technically implemented across interoperable agent frameworks. Enforcement actions against agentic deployments in financial services and healthcare will be an early indicator of whether regulators treat action-level observability as a baseline expectation or an advanced practice.