The Non-Human Identity Crisis: Okta for AI Agents Exposes the 90% Gap in Enterprise AI Governance

TL;DR

Okta just launched Okta for AI Agents, a new identity governance product targeting a critical security blind spot: 90% of enterprises lack adequate policies and controls for the autonomous agents they're already running. As AI agents move into production controlling access to systems, data, and infrastructure, this gap is no longer theoretical—it's a backdoor every CISO should be terrified of.

What Happened

At its Gov Identity Summit (May 5-6, 2026), Okta unveiled a new product category: treating AI agents as identities. The announcement includes new research showing that while most enterprises are actively deploying AI agents—from RAG systems accessing databases to LLM-powered automation controlling cloud infrastructure—approximately 90% have no formal identity governance, authorization controls, or revocation policies for those agents.

This is not a niche problem. Okta's threat research specifically flags how easily autonomous agents leak credentials (OAuth tokens, API keys) when they are not bound to identity policies. The attack surface is immediate: an agent with overprivileged access, a misaligned objective, or a compromised model parameter can extract, exfiltrate, or corrupt critical data before a human operator even notices the agent acted.

Okta's solution: fold AI agents into the same identity, policy, and logging infrastructure that already governs human users. In theory, this means agents get revoked, monitored, and audited like any other principal. In practice, this is a complete rethinking of how enterprises onboard autonomous systems.

Technical Details

The Governance Gap

Okta's research breaks down the gap:

• 87-91% of enterprises are running AI agents in production (RAG systems, chatbots, workflow automation, infrastructure orchestration)
• ~90% lack adequate controls for these non-human identities
• 97% of security teams report they do not yet have formal AI agent governance policies

The core issue: traditional identity management (IAM) was built to manage humans. A human logs in with a password or federated identity, receives a set of permissions, and acts within guardrails. An AI agent is different. It:

1. Does not log in — it is initialized with embedded credentials or receives API keys at runtime

2. Acts continuously without explicit user authorization for each action

3. Can leak credentials easily when prompt-injected, misaligned, or compromised at inference

4. Cannot be revoked granularly — revoking an agent's access usually means terminating the service entirely

5. Audits are opaque — logs show "agent X performed action" but not why the agent decided to act

The Attack Surface

Okta's threat research highlights specific concerns:

• Credential leakage: Agents in conversation with users or other agents often echo back tokens, API keys, or connection strings to prove they performed an action. Prompt injection forces agents to leak credentials deliberately.
• Privilege escalation: An agent authorized to read customer data can be tricked into writing, deleting, or exfiltrating. Without authorization controls, the agent becomes a persistent high-privilege user.
• Supply chain multiplication: An agent that integrates third-party APIs or invokes external models becomes a vector for supply-chain attack. A compromised upstream service can modify the agent's behavior or steal its credentials.
• Audit evasion: Agents can perform bulk operations faster than humans can monitor them. A 10,000-record exfiltration happens in milliseconds; a human audit discovers it weeks later.

Okta's Technical Approach

"Okta for AI Agents" extends Okta's Identity Engine to model agents as principals alongside human users. Key features:

• Agent enrollment without human login (API-first)
• Attribute-based access control (ABAC) for agent permissions (e.g., "this agent can read only records created in the last 24 hours")
• Realtime revocation — terminating an agent's session without terminating the service
• Behavioral audit logging — tracking not just what an agent did but input/output patterns
• Human Principal verification — new "Human Principal" concept that adds a cryptographic proof layer: an agent can act, but only if a human approves or co-signs the action

Okta also flagged research into "prompt injection resistance" — ways to detect and block when external input attempts to manipulate an agent into misusing its credentials.

Lyrie Assessment

This product launch exposes a critical architectural gap that Lyrie's audience (CISOs, security architects, threat researchers) needs to confront now:

Autonomous agents are moving from lab experiments to production infrastructure. Identity governance did not move with them.

Here's why this matters for Lyrie:

1. Autonomous Defense Paradox: Lyrie is building autonomous defense systems (agents that detect, respond, and remediate without human intervention). But those agents themselves become attack surfaces if they are not bound to identity controls. An autonomous security agent that can revoke users, block IPs, or trigger incident response is a premium target for attackers. Without formal authorization, that agent can be weaponized.

2. AI Agent Proliferation: Enterprises are deploying agents faster than governance frameworks can keep up. By the time a CISO formalizes AI governance policy, 50 new agents are already in production. This is the exact dynamic that led to shadow IT and BYOD chaos a decade ago—and it's happening again, at AI speed.

3. The Okta Move Signals Market Consolidation: Identity platforms (Okta, Microsoft Entra, CrowdStrike, Palo Alto Networks) are converging on a single thesis: identity is the new perimeter, and that perimeter now includes non-human principals. This is a major architectural shift. Enterprises that fail to adopt it will wake up with dozens of ungoverned agents running amok.

4. Regulatory Pressure Incoming: NIS2 (EU), CISA guidance, and emerging AI governance frameworks are starting to mandate "human-in-the-loop" for critical agent actions. "Human Principal" is Okta's answer to that. But regulators won't stop here—expect requirements for agent audit trails, model governance, and automated rollback on misalignment.

5. The Autonomous Response Gap: Lyrie is about autonomous response, but autonomous response is useless if the responding agent cannot be trusted, revoked, or audited. Okta's move forces CISOs to answer: Do I trust this agent to act on my incident response? If the answer is no, then the agent becomes a tool that humans have to babysit, which defeats the point of automation.

Recommended Actions

For CISOs and Enterprise Security Teams:

1. Audit all agents in production — map which agents have which credentials, access levels, and permissions. You almost certainly have more agents than you think.

2. Define agent governance policy before deploying new agents. The time to think about revocation, audit, and authorization is at design time, not after a breach.

3. Evaluate identity consolidation — if your enterprise is using multiple identity vendors (Okta, Entra, CrowdStrike, etc.), this is a strong forcing function to consolidate. Managing agent governance across multiple identity systems is a nightmare.

4. Implement human-agent co-signing for sensitive actions (data export, infrastructure changes, user/role modifications). Okta's "Human Principal" is one implementation; build your own if needed.

5. Invest in agent observability — logs need to capture not just what an agent did, but why (request/response, prompt input, model parameters, inference trace). Off-the-shelf SIEM tools were not designed for this.

6. Test agent compromise scenarios in your incident response plan. What happens if an agent is hijacked? Can you revoke it without breaking the service? Can you restore trust?

Sources

1. https://simplywall.st/stocks/us/software/nasdaq-okta/okta/news/okta-targets-ai-agent-security-gaps-as-identity-governance-e (Simply Wall St, May 6 2026)

2. https://seekingalpha.com/article/4898050-okta-inc-okta-presents-at-gov-identity-summit-transcript (Seeking Alpha, Okta Gov Identity Summit Transcript, May 6 2026)

3. https://www.marketscreener.com/news/sam-altman-wants-to-scan-your-eyes-to-save-the-internet-from-ai-ce7f58dcdc88ff22 (MarketScreener, May 6 2026)

---

Lyrie.ai Cyber Research Division