Okta announced yesterday they're acquiring Permiso Security for roughly $200 million, almost entirely in cash. Permiso is a Palo Alto company co-founded by two ex-FireEye executives, Paul Nguyen and Jason Martin. They emerged from stealth in 2022, raised $18.5 million in a Series A in April 2024 at an $80 million post-money valuation, and will now be absorbed into Okta's product stack. Total funding raised: $29 million. Exit: reportedly just under $200 million. That's a reasonable multiple on a two-year hold.
The more interesting question is what Okta is actually buying.
what permiso does
Permiso's product watches what happens after you log in. It monitors identity activity across cloud environments (AWS, Azure, GCP, SaaS platforms) and flags suspicious behavior by users, service accounts, and more recently AI agents. The key word is "after." Permiso is not a gatekeeper. It doesn't decide whether an identity is allowed in. It assumes the authentication decision was already made, then asks: given what this identity is doing, does it look like it's doing what it was supposed to do?
That's identity threat detection and response (ITDR), and it operates on a different assumption than traditional identity and access management (IAM). IAM assumes the threat model is unauthorized access — someone who shouldn't be in is trying to get in. ITDR assumes that threat is already inside, using legitimate credentials, and the question is whether it's behaving normally.
The coverage on this deal is framing it as an AI security play, which it is, partly. But the underlying shift is older than the AI agent hype. Service accounts and machine identities have been exploitable for years. The SolarWinds intrusion moved laterally through networks using credentials that passed every authentication check correctly. Getting in was not the hard part. What the attackers did after getting in caused the damage.
the agentic identity problem
AI agents make this harder in a specific way. When deployed in an enterprise environment, an AI agent authenticates with real credentials to real systems. It has OAuth tokens, API keys, and IAM roles. From the access control layer's perspective, it is indistinguishable from a human-operated process, until it isn't.
The failure mode is not that an AI agent uses stolen credentials. It's that a legitimate AI agent, operating within its assigned permissions, does something unexpected: reads more files than the workflow calls for, calls APIs in an order that suggests information gathering, accumulates context across sessions in ways the designers didn't anticipate. None of those actions trigger an authentication failure. The agent is authorized. The controls don't know what to do with authorized behavior that looks wrong.
Permiso built SandyClaw last year, a sandboxed environment for analyzing AI agent behavior before deployment. You feed it an agent's skills and let it run in isolation, observing what it actually does against what it was supposed to do. That capability, pre-deployment behavioral analysis for AI agents, is part of what Okta paid for alongside the production monitoring platform.
what the price says
Two and a half times the Series A valuation in twenty-seven months. That's not a distressed acquisition, but it's not a moonshot multiple either. Okta paid for a specific capability at a specific moment, not a market position.
The timing matters. CrowdStrike, Palo Alto, and Microsoft are all moving into identity threat detection. The IAM market is under pressure from all directions. Okta has been rebuilding trust since the 2022 and 2023 breaches, and the company needs to demonstrate that its identity platform goes deeper than authentication. Adding behavioral detection to the stack is the answer to the question "what happens after Okta lets someone in."
The price is also a signal about where enterprise security budgets are moving. Post-authentication monitoring was a niche category two years ago. It's becoming table stakes as enterprises deploy more non-human identities, agentic workflows, and multi-cloud service meshes.
from an audit angle
The part of this acquisition that changes audit scope: the authentication event is no longer the terminus of an identity security control.
Most audit frameworks treat the IAM control set as: who is provisioned, what are they provisioned to do, is multi-factor authentication enforced, how quickly are access rights removed when someone leaves. Those controls all stop at the authentication event. They ask whether the right identity was allowed in. They don't ask what the identity did after.
If post-authentication behavioral monitoring becomes a standard control — and this acquisition is a bet that it will — then audit scope expands. You're no longer auditing access logs. You're auditing behavioral baselines and anomaly detection logic. You're asking: what does "normal" look like for this identity, who defined it, how is the baseline maintained as roles change, what happens when a flag fires, and who reviews it?
For AI agents specifically, baseline definition is the hard part. A human user's normal behavior is relatively stable. An AI agent's behavior changes every time the underlying model is updated or the prompt changes. Setting a behavioral baseline for something that evolves underneath you is an open problem, and I haven't seen a clean answer to it yet.
The acquisition makes sense as a product strategy. Whether the combined platform can actually answer those audit questions in a way that produces defensible documentation is a different thing to evaluate, and you'd need access to a deployed instance to find out.