GitGuardian / NHIcon 2026 — Agentic AI forces a paradigm shift in non-human identity security
• Category: Security
AI relevance: AI agents autonomously traverse APIs, data stores, and internal tools without human oversight — traditional IAM built for session-based human workflows cannot govern their dynamic, goal-seeking behavior, creating an exponentially growing class of unauditable non-human identities.
- The second annual NHIcon conference (hosted by Aembit) brought together identity security leaders from GitGuardian, Cloud Security Alliance, Snowflake, and academia to address why existing IAM is structurally inadequate for agentic AI systems.
- Aembit CEO David Goldschlag proposed three pillars for securing agentic systems: identity (who the agent is), invocation context (why it's acting), and secretless execution (ephemeral, just-in-time credentials per action). Zero Trust for agents must blend the agent's identity with the delegating human's context.
- AI security researcher Ken Huang introduced a two-class taxonomy: persistent agents (long-lived, stateful across sessions) and ephemeral agents (short-lived, task-scoped with distinct identity per execution). Both classes should be backed by verifiable credentials anchored in decentralized identifiers (DIDs) — not long-lived tokens with no semantic meaning.
- CSA's John Yeoh framed the problem at scale: every tech shift (web → cloud → mobile → agents) created identity proliferation, but agentic AI does it exponentially. Agents spawn sub-agents, call other agents, and transfer data — creating a dynamic web of identities that defy static policies. Prompt injection becomes a vector for NHIs to seize credentials and APIs.
- Snowflake's Gaurav Singodia introduced the concept of "identity drift" — permissions that made sense at agent creation may no longer align with the agent's evolved behavior. Unlike scripts that execute predetermined instructions, agents pursue goals and adapt, meaning their effective privilege scope changes over time without any policy update.
- Key consensus: the industry must shift from point-in-time access checks to continuous behavioral validation — monitoring an agent's actions against its stated purpose in real time, not retrospectively auditing after damage is done.
- Traditional controls (static roles, session tokens, long-lived credentials) actively magnify risk in agentic contexts because they were designed for deterministic, human-driven workflows that return the same result from the same input.
Why it matters
- This conference represents the first coordinated industry effort to define identity frameworks specifically for AI agents — moving past "agents are ungoverned" problem statements into concrete architectural proposals (secretless execution, DID-anchored credentials, continuous validation).
- The "identity drift" concept fills a gap in current security thinking: most agent governance focuses on deployment-time permissions, but agents evolve their behavior post-deployment. This means static RBAC is fundamentally broken for agentic systems.
- The persistent-vs-ephemeral agent taxonomy has direct implications for infrastructure teams: persistent agents need credential rotation and behavioral baselines; ephemeral agents need per-invocation identity issuance at scale — radically different plumbing than traditional service accounts.
What to do
- Eliminate long-lived agent credentials: replace static API keys and service account tokens with ephemeral, cryptographically anchored credentials scoped to each action or task.
- Implement invocation context: every agent action should carry metadata about who delegated the task, why the agent is acting, and what it's authorized to do at that specific moment.
- Monitor for identity drift: establish behavioral baselines for deployed agents and alert when an agent's actual access patterns diverge from its stated purpose — don't wait for incident response.
- Classify your agents: distinguish persistent agents (need state management, credential rotation, behavioral monitoring) from ephemeral agents (need per-execution identity issuance, automatic credential expiry).
- Watch the NHIcon recordings: the full conference is available on Aembit's site and covers implementation details beyond what summaries can capture.