Wiz's August 17, 2026 report describes a GitHub Actions injection in a Snowflake repository and research access to internal Jira data. Its update clarifies that Copilot checked the merged pull request and missed the vulnerability; whether the vulnerable change was AI-assisted is unclear. The report says Snowflake remediated the issue and rotated the affected credential. Wiz's report and correction.
Distinguish authorship, review and execution
An AI tool participating in a pull request does not establish who introduced a specific line. Preserve that distinction when investigating an incident or choosing controls. A review that returned no finding is not proof that a workflow cannot be exploited.
Check what the workflow can do
For a firm approving an implementation, ask what untrusted input reaches the workflow, which credentials it can use and what an attacker could change. Verify the restrictions outside the model's narrative explanation.
Test before and after changes
Use representative permitted and prohibited cases. Check that a rejected request produces no unintended write, message or external action. Retain the evidence needed to reconstruct the result and assign someone to review material changes.
These are recommended checks for a scoped implementation. The incident does not establish an error rate for law-firm systems or prove that AI-authored code caused this vulnerability.
Discuss implementation requirements.
Editorial correction: unsupported or unavailable-source claims were revised. Original publication date retained; this correction does not certify every other article in the archive.