Sygnia Finds Authorization Flaw in Claude-Built Onboarding App

The application used temporary access tokens, expiration, rate limiting and audit logging. Those controls covered what happened after a token was issued, but the issuance and restoration flow treated possession of an applicant's globally unique identifier, or GUID, as sufficient proof that the requester was entitled to that applicant's record.

Sygnia said the flaw could let a low-privilege user obtain access to another applicant's information, including identity-verification and payment data. The security firm said an LLM used during the penetration test surfaced the issue. The public disclosures do not name the financial institution, and they do not report a confirmed real-world breach or exploitation; the finding comes from an authorized assessment.

A GUID identifies a record, but it is not evidence that the requester owns or controls that record. Token expiry, rate limits and logs can reduce exposure or aid investigation, yet none substitutes for an authorization decision before the token is created or restored.

The relevant test is therefore not only whether a valid applicant can recover access. Reviewers also need negative tests showing that one applicant cannot obtain another applicant's token, even after discovering or guessing a record identifier.

For teams reviewing AI-generated authentication flows, the useful question is where entitlement is proved. Each token-issuance and recovery path should bind the requester to the target account through an independent control, enforce authorization server-side and test cross-account failures explicitly.

This is a single, anonymized case disclosed by the security firm that performed the assessment, so it does not establish a failure rate for Claude-generated code. It does illustrate a broader review risk: generated code can contain familiar security mechanisms while omitting the business-rule check that makes those mechanisms meaningful.