Every AI governance conversation in 2026 eventually arrives at the same uncomfortable admission: most companies can tell you what their agents did, but not what they are about to do.

JetStream Security is trying to change that. On September 2, the startup — backed by $34 million in seed funding from Redpoint Ventures and the CrowdStrike Falcon Fund — announced Clearance, a reasoning engine that evaluates each AI agent action before it executes and decides whether it should be allowed to run.

The difference matters more than it sounds. Traditional AI gateways sit in the request path and route or monitor traffic. Clearance does something different: it reads the agent’s intended next step against an approved design, decides yes or no in real time, and blocks anything that falls outside what the agent is supposed to do — including dangerous sequences of actions that each look fine individually but are harmful in combination.

What the Numbers Show

JetStream built Clearance against a backdrop of research that keeps landing in the same place. A survey by Enterprise Management Associates and Cequence Security found that 65 percent of enterprises have seen their AI agents act beyond their intended scope, with 29 percent reporting measurable organizational impact. Only 34 percent of respondents said their organisation evaluates whether an agent is actually authorized to act at execution time. The rest rely on standing permissions, periodic reviews, or inherited access — none of which stop an agent mid-sequence.

JetStream’s founders — drawn from CrowdStrike, SentinelOne, Dazz, Cohesity, McAfee, and Attivo Networks — spent years watching post-hoc detection fail in traditional endpoint security. Clearance applies the same lesson to AI: once an action has executed, the data is gone, the API is called, the file is written. Prevention is cheaper than remediation.

How Clearance Works

At the foundation is a concept JetStream calls AI Blueprints: dynamic graphs that map how AI systems actually operate — which agents use which models, which data they can touch, which tools are in scope, and which identities or API keys are involved in each step. Policy binds to identity when access is requested, and Clearance reaches inside each tool call down to the individual parameter. A team can allow an agent to read a CRM record while blocking it from updating or deleting one. The same tool, different permission, different outcome.

The reasoning happens in flight, ahead of execution. JetStream’s term for this is pre-execution authorization, and it closes the gap that post-hoc detection tools leave open: the window between the moment an agent decides what to do and the moment it does it.

Clearance also tracks the cost of AI operations in real time — showing how much individual agents cost to run and which teams or owners are responsible. In regulated industries and in any organisation running AI at scale, that visibility is increasingly required. Auditors want to know what an agent did over the last 30 days. Currently, just 46 percent of enterprises say they could produce a complete audit trail. Clearance creates that trail automatically.

The product demonstrated at Fal.Con in Las Vegas from August 31 to September 2 and enters general availability this fall.

What This Means for Business

If you have deployed AI agents across any part of your business — customer service, finance operations, data analytics, internal reporting — you are probably in the 65 percent. Your agents may be behaving exactly as designed, or they may be accessing data they should not, calling tools outside their remit, or performing sequences of actions that no single policy review caught.

The response so far has mostly been logging. Companies instrument their agent calls, ingest the events into SIEMs, and alert on anomalies after the fact. JetStream’s argument is that this is the wrong model for agentic workloads, where a chain of five authorized-looking steps can result in an unauthorized outcome.

The broader point is about trust architecture. Agents in 2026 operate across multiple systems — your CRM, your data warehouse, your internal knowledge base, your communication tools. Giving each of them a broad access token and checking whether anything went wrong afterward is a pattern that worked poorly in traditional security and works even less well when the actor is autonomous and fast.

There is a parallel story from EDNA’s own work with Omni clients. The companies that deploy AI agent workforces successfully tend to start with tightly scoped agents — one job, one dataset, one output. Clearance formalizes that discipline at the infrastructure level, which is the right direction. Narrow agents with verified permissions are more reliable, easier to audit, and easier to explain to the business units they serve.

The market for AI governance tooling is growing fast. What JetStream is adding is the execution-time layer — the piece that sits between an agent’s intent and its action. Whether or not Clearance becomes the standard, the category it represents will.

---

JetStream Clearance enters general availability in fall 2026. For businesses deploying AI agent workforces, EDNA’s Omni Ops service includes design principles for scoped, auditable agent deployment. Learn more about Omni Ops.