AI Pentest
The AI feature your customers use: what it reads, what it reveals and which tools it can call on someone else's say-so.
Compliance / SOC 2 / CC4.1
SOC 2 does not strictly require a penetration test. A point of focus under criterion CC4.1 names penetration testing among the separate evaluations a company can use to check that its controls work. This page shows where it sits in the criteria, what they leave to you, and what a useful report contains.
TSC / CC4 MONITORING ACTIVITIES / CC4.1 / POINT OF FOCUSNo. A SOC 2 report tests controls against the AICPA Trust Services Criteria, and the criteria come with points of focus: examples of what a control may address. The criteria state that using them does not require an assessment of whether each point of focus is addressed.Note 4
Penetration testing appears in one of those points of focus, under CC4.1, among the ongoing and separate evaluations an organisation can use to check that its controls are present and functioning.Note 5
So the test is optional on paper. In practice your SOC 2 report is read by your customers' security teams, and the controls it covers, from logical access to change management, are the ones a penetration test tries to get around. A recent test shows that the controls work, not only that they exist.
The criteria leave scope to you, so tie it to the system your SOC 2 report describes. For most SaaS companies a useful test covers:
Leave out what the report leaves out, and say so. A test wider than the system description is fine; a test narrower than it leaves a gap someone will ask about.
Usually, yes. If you hold or pursue ISO/IEC 27001:2022, a pentest report is common evidence for two controls in its Annex A: 8.8, management of technical vulnerabilities, and 8.29, security testing in development and acceptance.Note 8
As those reproductions read, neither control names a penetration test; a test report is one way to evidence both.Note 9 Check with your certification body which evidence it expects for your scope.
For most SaaS companies that is the platform; for a growing number, also the AI feature on it.
The AI feature your customers use: what it reads, what it reveals and which tools it can call on someone else's say-so.
The application, its APIs and the cloud account behind them: authorization between tenants, authentication, sessions and configuration.
The SaaS platform and its AI feature in one scope and one report, including the paths that run from the conversation into the stack.
Once in each audit period, timed so the report and the retest both land inside it, and again after a major change to the system, such as a new product area, a new cloud account or an AI feature. That is our advice, not a rule from the criteria.
No. Scanning is a recurring control that finds known weaknesses; a pentest is a separate evaluation that shows what an attacker can do with them, including business logic and authorization flaws a scanner does not model. The CC4.1 point of focus names both kinds of evaluation.
Fix it and have it retested. A finding that was found, fixed and confirmed fixed inside the period is evidence that your monitoring works. Hiding it is the only outcome that looks bad, and the report is the record that shows you did not.
Yes, when it is part of the system your report describes. Launch Clearance tests the SaaS platform and its AI feature together, in one scope with one report, including the paths that run from the conversation into the stack, so your auditor reads one document.
Scope, method, findings with CVSS v4.0, retest status and the attestation letter.
Published ranges per package, excluding VAT.
Why their supply chain duty reaches your questionnaire.