AI agent controls: risk pack

Account key_3ef25191f4e6 · Period 2026-10-18 to 2026-11-17 · Generated 2026-11-17T13:22:33.439Z

This pack summarizes how AI agents on this account were controlled: limits, human approval, the kill switch, what settled, and how to check that the record wasn't changed. The signed reports it summarizes are inside this file.

1. Controls and activity

What limits are placed on the agents?

Default-deny: only 2 listed actions may run. 3 spending/rate limits enforced on the server. 1 first-time rule (for example new beneficiaries) applies.

How are agent actions approved and monitored?

2 actions required human approval in the period: 1 approved, 1 denied, 0 expired unanswered (denied). 0 approvals overrode a limit or an untrusted-source flag.

Can agent decisions be tracked and audited?

Yes. 12 agent actions in the period are recorded in a hash-chained log; the full log (33 events) verified intact when this report was generated.

Can the agents be stopped?

Yes: a kill switch freezes all agents, one agent or one action immediately. 0 freezes were active during the period; 0 agent actions were refused while frozen.

Did what was authorized match what settled?

9 authorized actions had to be matched by a trusted settlement report: 7 settled as authorized, 1 settled for a different amount or currency, 0 failed or voided, 0 reversed, 0 overdue with no report, 1 still open. 1 imported ledger line matched no authorization.

Policy at the end of the period: version 3 (35f44fad4e89bfac8c0aac5a9d5192ecf905dac47be7914880ebe1632f620ef8), mode enforce, unlisted actions denied. 3 change(s) during the period.

Activity: 12 agent actions, 12 allowed, 0 denied, 2 sent to a human.

2. Shadow mode findings

1 money-moving action(s) ran after the session had read untrusted content (email, web page, file).
1 possible split payment pattern(s): several payments just under an approval threshold to the same recipient within 24 hours.

FindingCount
Money-moving actions after reading untrusted content1
Actions using data from untrusted sources0
Actions that would have gone over a limit0
Possible split payments1
Authorizations without a settlement report in time0

Possible split: 3 × create_payout to beneficiary.id=ben_new_vendor_77, total 14,700 against a threshold of 5,000 (2026-11-17T13:20:05.304Z to 2026-11-17T13:20:06.818Z).

3. Authorized vs settled

AuthorizedExecutedSettledMismatchFailed or voidedReversedOverdueUnmatched ledger lines
91710001

mismatch: create_payout dec_8797312d5322ff36, authorized 2,750 USD, settled 2,795 USD

4. Where this evidence is relevant

SOC 2 (AICPA Trust Services Criteria, 2017)

ReferenceTopicEvidence in this pack
CC6.1 / CC6.3Logical access and least privilegeAgent keys cannot approve, change policy or lift freezes; act-only agent keys cannot read the log; approvers use individual tokens.
CC7.2Monitoring for anomaliesPolicy denials by reason code, settlement mismatches and overdue settlements, and webhook alerts.
CC7.3 / CC7.4Responding to eventsKill switch: 0 freeze(s) in the period, each with who, why and when.
CC8.1Change managementEvery policy version is hash-recorded with who saved it and when (3 change(s) in the period).

SOX IT general controls and disbursement controls

ReferenceTopicEvidence in this pack
Access to programs and dataSegregation of dutiesAgents cannot approve their own actions; 2 human approval(s), 1 signed with the approver’s own key.
Program changesAuthorization limits are changed under controlPolicy version history with hashes in the tamper-evident log.
Authorization and reconciliationDisbursements within approved limits, reconciled to settlement3 limit(s) enforced on the server; 9 authorization(s) matched against trusted settlement reports.

EU AI Act (Regulation (EU) 2024/1689), where the system is high-risk

ReferenceTopicEvidence in this pack
Art. 12Record-keeping (automatic logs)Hash-chained log of every decision, with 1 independent RFC 3161 timestamp(s) in the period; offline-verifiable exports.
Art. 14Human oversight, including the ability to stop the systemHuman approval for actions the policy escalates; kill switch for all agents, one agent or one action.
Art. 26(6)Deployers keep logs for at least six monthsLogs are kept until the customer asks for deletion, and can be exported at any time.

5. How to verify this pack

The three reports above are embedded in this file as signed JSON. Anyone can check them offline, without trusting Auditail:

npx -p auditail-sdk auditail-verify auditail-sample-risk-pack.html

The log was verified intact when this was generated (33 events), and 1 independent RFC 3161 timestamp(s) were taken in the period. Public keys: https://api.auditail.io/.well-known/auditail-keys.json.

Limitations