The dashcam
for AI agents
Independent, sealed evidence of what your agents did, verifiable without relying on the platform that acted.
Evidence, not assertions.
Platforms run AI agents. Guardrails keep them in check.VIAS keeps an independent record of what they did.
01 · Runtime
Executes
Model providers, frameworks and agent orchestrators.
What is executing, and under whose authority?
02 · Control
Constrains
Sandboxes, guardrails, policy engines and hardware watchdogs.
Can we contain unauthorised behaviour?
03 · Evidence — VIAS
Seals
Out-of-band sealing of submitted records, manifest binding and coverage-gap analysis.
Can a reviewer verify the record and its stated coverage?
04 · Liability & review
Relies on it
GRC teams, auditors, counsel, underwriters and boards.
What can be established when an action is disputed?
A dashcam doesn't prevent the collision. It establishes what happened.
Who needs it.
Security incidents
What did the agent attempt, and what controls intervened?
Claims and underwriting
What record existed before liability arose?
Payments
Did execution remain within the authority granted?
Administrative decisions
What inputs, authority and process produced the outcome?
Audit and governance
Can evidence be verified without relying solely on the operator?
From submission to offline verification in five steps.
01
Submit
The operator, or its observability stack, sends an execution manifest to VIAS, which seals what it receives. No runtime re-architecture.
02
Attest
The Governance Gate applies a rules-based comparison against the registered authority grant — not a judgement of whether the action was right.
03
Seal
The record is signed, sequenced and linked to the previous record. Fields that were not submitted are labelled as gaps.
04
Export
Sealed records and their coverage statement are exported as a self-contained evidence pack.
05
Verify
The evidence pack is designed so a reviewer can check signature, sequence and vectors against the published public key, offline, without a live VIAS service. The signed-export verification chain is pending end-to-end testing.
Every record is assessed across six dimensions (input, origin, response, handoff, scope and authorisation), with anything not submitted labelled as a gap. See how it works →
Why separation from the actor matters
A system's internal logs remain valuable operational telemetry. But when its actions are challenged, investigators may also need evidence sealed outside the decision environment. VIAS provides that separate evidence layer.
Agent Platform Logs
VIAS evidence layer
Location
Purpose
Record handling
Verification
Intended use
Out-of-band from the agent is not the same as independent of the vendor
Hardware containment tools
Hardware containment tools from the leading chip vendor are control-plane measures — valuable, and increasingly required. But they run on the vendor's own silicon and are verified by the vendor's own tooling. They are out-of-band from the agent, but not independent of the vendor.
VIAS evidence custody
VIAS evidence is sealed independently of both the agent runtime and the underlying compute vendor. The custody boundary is the actor, not the infrastructure provider — so the evidence survives even if the vendor's own tooling is the thing later disputed.
Evidentiary separation from the actor is one boundary. Vendor-independence from the infrastructure provider is another. VIAS holds both by design.
Why custody comes before the dispute.
Most records are gathered after an incident.
VIAS seals submitted records before a claim, investigation or dispute exists.
The evidentiary value comes from preservation before incentives to alter the record emerge.
What's preserved doesn't have to be what's shared. You decide what's disclosed, to whom and when.
As AI agents take on consequential work, the question is no longer whether to keep a record, but who controls it.