Attestation
JumpCloud's "Proof Gap": Most Firms Can't Show What Their Agents Actually Did
A new JumpCloud report, published October 7, puts a number on something security teams have been saying anecdotally for months: knowing what an AI agent is allowed to do and proving what it actually did are two different capabilities, and most organizations only have the first.
The report found that 72% of organizations report a considerable gap between what AI is allowed to access and what they can prove it actually accessed. JumpCloud is calling this the "Proof Gap," and it sits alongside a related finding that more than one-third of surveyed organizations investigated an AI-related concern but could not determine what the AI had accessed.
This is a narrower and more specific claim than the usual "governance is behind" survey finding. It's not about whether agents have too much access (a separate, also-documented problem) — it's about evidentiary capability after the fact. An incident happens, someone asks what the agent touched, and in a third of cases the honest answer is: we don't know. That's not a policy gap, it's a logging and attribution gap, and it shows up again in a related JumpCloud data point: only 38% of organizations can fully identify the identity that authorized a given AI action, and only 36% can fully prove the AI had only the permissions it needed.
The company's pitch is unsurprising given it sells identity infrastructure — JumpCloud's senior VP for advanced technologies framed the fix around giving every agent clear ownership and lifecycle-long access controls. Take the vendor framing with appropriate skepticism, but the underlying measurement problem is real and distinct from access-scope complaints: even organizations that have scoped an agent's permissions tightly often can't reconstruct, after the fact, which specific actions it took under which authorization.
The practical implication is that audit trails for agentic systems need to be designed before an incident forces the question, not reconstructed after one. A 72% gap between "what's allowed" and "what's provable" means most deployed agent systems today would fail a basic forensic test: if an agent did something wrong, could the organization show, with evidence, exactly what happened and who authorized it? For most of the 250 IT leaders JumpCloud surveyed, right now, the answer is no.