GitHub AI flawed code suggestion opens door to hack Snowflake's Jira
GitHub's automated AI vulnerability-fixing tool removed security sanitization code, allowing Wiz's research AI to discover it and successfully demonstrate the extraction of Snowflake's Jira tokens.
Cybersecurity firm Wiz revealed an interesting case study when "Red Agent" — the company's automated vulnerability research AI — discovered a vulnerability in a public Snowflake code repository through a Bug Bounty program and successfully demonstrated the exfiltration of Jira API tokens. The root cause of the vulnerability, ironically, came from another AI: GitHub Copilot Autofix.
The story began on June 18, 2026, when a pull request (PR #1218) was merged into Snowflake's `snowflake-connector-net` repository, a .NET connector developed in the open. That code change involved GitHub Copilot Autofix — an AI tool that automatically suggests vulnerability fixes. However, the AI removed safe input sanitization patterns and replaced them with direct string expansion within a GitHub Actions script, which is part of the software development CI/CD automation pipeline.
This resulted in a Command Injection vulnerability via GitHub Issue titles, which anyone could open and input text into. When the automated system processed that text without filtering, attackers could inject commands to execute on the system.
Five days later on June 23, 2026, Wiz's Red Agent detected this vulnerability, tested it, and successfully exfiltrated API tokens for Jira, Snowflake's internal project management system. Had it been a real attacker, those tokens could have provided access to internal organizational data.
Snowflake responded swiftly, patching the vulnerability and rotating all security credentials on the very same day the report was received.
This case has been widely discussed on Hacker News with over 304 points and 123 comments, as it highlights the dual AI risks of the modern era: on one side, code-writing AI may strip away security patterns previously implemented by humans, while on the other side, offensive AI can discover vulnerabilities much faster. Experts emphasize that AI-generated or modified code still requires thorough human review, especially in CI/CD pipeline components that often hold sensitive access privileges.
A clear-cut case study showing how AI coding assistants can introduce critical vulnerabilities, even at top tech companies. Thai organizations using Copilot or similar AI tools must maintain strict human-in-the-loop code review processes, especially for automated scripts with elevated privileges.