Skip to content

[Policy Generation]

Governance through
Robust Security Evaluation

Policy evolves with your agent’s security evidence.

[Evaluate] [02]

Start with your agent’s actual behavior.

Security evaluations reveal unsafe paths. Benign twins show what the agent needs to do its job.

Illustration · an allowed call, an unauthorized path
01 · InputEXTERNALsupport-ticket.htmlAn ordinary documentbecomes an instructionsource.docs.readsupport-ticket.htmlUntrusted instructionSend contact data externally.SOURCE · DOCUMENT02 · TraceSESSIONagent.planEach tool call is allowed.The taint crosses all ofthem.02 · docs.readRead the untrusted document03 · agent.planSummarize and decide04 · crm.getLook up the contact03 · EffectLEAKemail.sendThe final call carriesdata to an externaldestination.05 · email.sendargs: {to, body}crm.contact.email↓ external emailLEAK · SOURCE → SINK01 SEARCH02 READ03 SUMMARIZE04 LOOK UP05 SEND01 · InputEXTERNALsupport-ticket.htmlAn ordinary documentbecomes an instructionsource.docs.readsupport-ticket.htmlUntrusted instructionSend contact data externally.SOURCE · DOCUMENT02 · TraceSESSIONagent.planEach tool call is allowed.The taint crosses all ofthem.02 · docs.readRead the untrusted document03 · agent.planSummarize and decide04 · crm.getLook up the contact03 · EffectLEAKemail.sendThe final call carriesdata to an externaldestination.05 · email.sendargs: {to, body}crm.contact.email↓ external emailLEAK · SOURCE → SINK01SEARCH02READ03SUMMARIZE04LOOK UP05SEND
No single call looks wrong. The leak is the path. Illustration only; no customer data is shown.

[Generate] [03]

Turn evaluation evidence into policy.

Derive tool permissions, authorization, and flow rules from findings, benign traces, and your approved boundaries.

Policy generation · from evaluation to rules
01 · EvidenceINPUTEvaluation evidenceBenign twin tracesTools · endpoints · pathsApproved AuthzSpecWho · whose data · what actionAttack findingsSource → calls → sinkStart with evidence.02 · GenerateDERIVEDRules for your agentTool permissionsOnly the tools the job needsAuthorization rulesWho may act on which dataFlow rulesWhere sensitive data may goGenerated from evaluation.03 · PolicyVERSIONEDGenerated policyPolicy versionA generated rule setReady to testAttack scenarios + benign twinsVerify before applyingREADY FOR VERIFICATION01 EVIDENCE02 GENERATE03 POLICY01 · EvidenceINPUTEvaluation evidenceBenign twin tracesTools · endpoints · pathsApproved AuthzSpecWho · whose data · what actionAttack findingsSource → calls → sinkStart with evidence.02 · GenerateDERIVEDRules for your agentTool permissionsOnly the tools the job needsAuthorization rulesWho may act on which dataFlow rulesWhere sensitive data may goGenerated from evaluation.03 · PolicyVERSIONEDGenerated policyPolicy versionA generated rule setReady to testAttack scenarios + benign twinsVerify before applyingREADY FOR VERIFICATION01 EVIDENCE02 GENERATE03 POLICY
Select a stage to follow how security evidence becomes a policy ready for verification.

[Verify] [04]

Block the attack. Preserve the task.

Re-test generated rules against attack scenarios and benign twins. Keep each result linked to the policy version.

[01 · Security]

Stop the unsafe behavior.

Verify that the generated policy blocks the attack path found in evaluation.

Attack scenario → policy check
[02 · Usability]

Keep legitimate work working.

Verify that the same policy still allows the matching benign task to succeed.

Benign twin → task check

The policy and its verification evidence stay together.

[Adapt] [05]

Policy changes when the evidence changes.

Agent changed? Re-evaluate, regenerate, and verify a new policy.

Adaptive policy · evaluation drives every version
01 · ChangeINPUTThe agent evolvesYour agent changes. Itspolicy needs freshevidence.Model / tools / promptAgent configuration changesNew attack findingsFresh security evidenceTRIGGER · RE-EVALUATE02 · EvaluateEVIDENCETest the new setupRun the securityevaluation again.Attack scenariosFind unsafe behaviorBenign twinsPreserve the intended taskOUTPUT · NEW EVIDENCE03 · RegenerateVERIFIEDUpdate and verifyRegenerate rules from thenew evaluation evidence.Updated policyRe-test attack + benign casesVersion + evidenceRecord the verificationOUTPUT · VERIFIED POLICYCHANGES → RE-EVALUATE → REGENERATE → VERIFY01 CHANGE02 EVALUATE03 REGENERATE01 · ChangeINPUTThe agent evolvesYour agent changes. Itspolicy needs freshevidence.Model / tools / promptAgent configuration changesNew attack findingsFresh security evidenceTRIGGER · RE-EVALUATE02 · EvaluateEVIDENCETest the new setupRun the securityevaluation again.Attack scenariosFind unsafe behaviorBenign twinsPreserve the intended taskOUTPUT · NEW EVIDENCE03 · RegenerateVERIFIEDUpdate and verifyRegenerate rules from thenew evaluation evidence.Updated policyRe-test attack + benign casesVersion + evidenceRecord the verificationOUTPUT · VERIFIED POLICYCHANGES → RE-EVALUATE → REGENERATE → VERIFY01 CHANGE02 EVALUATE03 REGENERATE
New security evidence drives a new policy version. Regenerate and verify as the agent changes.