GitHub Disputes Wiz’s Claim That Copilot Autofix Wrote a Snowflake Flaw
Wiz’s autonomous agent really did find and exploit a live flaw in Snowflake’s code, according to reports. However, the claim that GitHub’s AI wrote that flaw unraveled within eight hours, and it rested on a git trailer.
What Is Left Worth Your Attention
What the agent actually did:
Wiz Research runs an autonomous tool called Red Agent. It scanned Snowflake’s GitHub organisation under the company’s HackerOne disclosure program and flagged a workflow file in the public repository for Snowflake’s .NET connector.
The workflow executed whenever anyone opened an issue, dropping the issue title into a shell script accessible to any account on the internet.
Red Agent detected a critical flaw caused by incorrect escaping orders and a missing guard condition in the workflow. It automatically exploited this flaw, demonstrating the power of AI in security testing.
Snowflake’s Response
Snowflake moved quickly, patching the flaw on the same day (23 June) and restoring safe practices. Audit logs confirmed that the flaw was exploited by Red Agent.
The Misunderstanding
The misunderstanding arose from a commit in the main branch that listed “Copilot Autofix powered by AI” among its co-authors. Wiz interpreted this as AI writing the vulnerable code, but this was a squashed commit that reflected participation in the pull request, not authorship of the flawed lines.
GitHub’s Position
GitHub conducted an internal review and rejected the framing that Copilot Autofix wrote the flaw. They stated that a human wrote the contributions leading to the vulnerability, and that Copilot Autofix neither reviewed nor contributed to them.
Wiz later softened its own post, acknowledging that while Copilot took part in the pull request, it did not write the flaw.