Who is this agent? What is it allowed to touch? Can you stop it right now? What did it do? Candor Agent Defense is agent cybersecurity at runtime: the controls and the dated evidence to answer, so the deal is not lost in the review.
Observe first on your staging fleet. Nothing is blocked until you have read the report.
Channel and Zoe, the Candor mascots. The four questions are drawn from 2026 enterprise questionnaires for agent vendors.
Candor does not take over your obligations to your customer. It provides the controls and the evidence that help you meet them, inside the product you already ship.
The same two planes Candor runs on its own fleet, deployed where you need them.
Evidence that travels with your product into every security review.
Start on your staging fleet. Then let Candor join the call.
At every governed action, Agent Defense asks who is acting, what they were granted, and whether the action itself is safe. Four illustrative outcomes.
Choose a scenario.
The sample policy requires a named approval before funds move.
Local illustrations with sample policies and amounts. They do not run an agent, move money or connect to a live Candor service.
Candor-reported results from its own test range and fleet, by date. Specific scenarios, not an independent certification and not a guarantee for your fleet.
A revoked agent could not regain access by old, forged, borrowed or no credential. Scope and unapproved destinations refused.
Requests in flight after a terminate were all stopped within 1.4 seconds; the next attempt was refused.
A compromised supervisor drill revoked all ten agents on both planes. Replacements healthy in under three minutes.
No degradation. 90th-percentile decision latency 2.4 seconds on the sovereign local model.
Identical outcomes in Candor's native environment and an NVIDIA OpenShell sandbox.
Each figure is a separate observation. Your fleet, integration and covered action paths are measured during Discovery, and the report states where coverage ends.
Read the evidence record ↗If Candor runs inside your product, it appears in your customer's vendor review. Two answers, both already true in the architecture.
Engine and authority plane run in your environment or your customer's. The safety model is Candor's own local model, and egress is sealed. Your customer's data does not have to reach Candor.
Your evidence pack includes Candor's dated tests, so your customer's reviewer does not start from zero. Mapped to NIST AI RMF, OWASP ASI and CSA AI-CAIQ, stated as mappings, not certifications.
Your contract is with your customer. Candor provides the controls and the evidence that help you meet it. The report says what it covers and what it does not.
Discovery observes without blocking. You receive an inventory, findings from both planes, a gap list against the questionnaire and a signed, chain-verified report that states what would have been contained and where coverage ends.
5 business days
10 business days
15 business days
Scoped jointly
Prices in USD, exclusive of taxes. The Discovery fee is credited in full toward implementation of the matching tier if you proceed within 30 days of the final report. Scope and terms are confirmed before work begins.
Full engagement details ↗A useful security conversation includes the controls, the evidence, and the parts still to be tested in your environment.
Not necessarily. Engine and authority plane can run in your environment or your customer's, with Candor's local safety model and sealed egress, so customer data does not have to reach Candor. The deployment model is agreed during Discovery and stated in the report.
Termination revokes the agent's authority first, so a cached, replayed or in-flight request is refused wherever it turns up, and then the execution context is stopped. On Candor's Fly range on 6 October 2026, requests in flight after a terminate were all stopped within 1.4 seconds. Your fleet is measured during Discovery.
A named person, through the approval route agreed during onboarding. The approval is recorded and usable once. An agent cannot approve its own high-impact action.
Candor reported a 90th-percentile decision latency of 2.4 seconds across 600 consecutive evaluations on its local model on 15 September 2026, on its test range. Your integration and operating mode need their own measurement.
Candor's evidence is mapped to NIST AI RMF, OWASP ASI and CSA AI-CAIQ. Mapping is not certification, and the report says so. Certification status is stated plainly in the engagement, never implied.
When one agent asks another to act, the acting agent's own authority governs. The requester's narrower scope is not yet evaluated by the authority plane.
The report and its manifest exist. A scoped, read-only credential for your customer's audit team is not yet available.
An action path that reaches a protected resource without passing through the gateway sits outside coverage. The deployment identifies those paths and the report states the boundary.