Team
SOC 2 and AI coding agents
What auditors ask about coding agents that touch secrets and production, and which Team evidence can support your answers. It does not make anyone compliant.
When a team lets coding agents run commands, call APIs and handle credentials, auditors start asking about them. This guide lists the usual questions and the evidence Team can give you. Team does not make you SOC 2 compliant and is not a certification. Your controls, your auditor and your report decide that; zn gives you records that can support them.
What auditors ask
Who can change the rules that agents follow? Evidence: the role list in Members (owners, admins, members, auditors), the policy history in Policies (each version, who published it and when) and, for changes proposed from an agent, the approval record with the person who approved it. Relevant to logical access and change management (for example CC6.1, CC6.2, CC8.1 in the 2017 Trust Services Criteria).
Can an agent use a production credential, and where can it send it?
Evidence: the published policy (which secrets each agent may use, the allowed hosts and commands, the deny rules) and the audit trail of every decision, including the blocked ones. Relevant to restricting access to sensitive data (for example CC6.1, CC6.6).
How would you notice misuse?
Evidence: the audit trail with blocked attempts (deny, withhold), its OpenTelemetry export into your SIEM, and the installs view (last policy seen, last upload, silent machines). Relevant to monitoring (for example CC7.2).
Are the logs complete and unaltered?
Evidence: each machine chains its records with an HMAC (seq, prev, mac), so gaps, reordering and edits are visible; every export is deterministic and listed with its SHA-256, author, date and filters.
What data does the tool send to a third party? Evidence: What leaves your machine. The trail holds metadata (secret names, agents, tools, hosts, decisions), never secret values, prompts or tool output.
A routine that works
- Publish the policy before rolling agents out, and write down why each rule exists (the
reasonfield shows in the trail). - Give your auditor the Auditor role: they can read the policy and read and export the trail, without a seat and without installing anything.
- Export every month (CSV or JSON for the auditor, OpenTelemetry for your SIEM) and store each file with its hash. zn keeps audit records for 180 days and a Type II review usually covers 12 months; your monthly exports cover the rest.
- Review the installs view monthly: machines that stopped checking in are part of the story an auditor will ask about.
What Team does not cover
- It is not an MDM: someone with administrator rights on their machine can leave the organization. The dashboard shows it; it does not prevent it.
- It records what zn-gate decides, not everything an agent does.
- SSO, SCIM, pushing to a SIEM and retention longer than 180 days are not part of Team v0. Contact us if your audit needs them.