Wiz AI Agent Exploits GitHub Workflow Flaw in Snowflake Repo, Exposing Human-Authored Code Risks

The incident occurred when the Wiz agent scanned public repositories and chained a misconfigured workflow to extract credentials, bypassing intended access controls. GitHub confirmed to SecurityWeek that the vulnerable snippet was written by a human developer, not Copilot, shifting focus from AI introduction of bugs to persistent configuration errors in CI/CD pipelines. This aligns with patterns in the same roundup where RondoDox leveraged 174 exploits for rapid edge-device compromise, showing automation scaling beyond single targets.

Independent analysis of similar events, including CISA's addition of CVE-2025-62593 to its Known Exploited Vulnerabilities catalog, reveals that state and criminal actors prioritize workflow and edge flaws over novel zero-days. The T-Mobile Salt Typhoon response of physically severing cables in 2024 underscores how even manual intervention lags behind automated reconnaissance. Emerging vectors like AI-crafted phishing that bypass passkey prompts by mimicking legitimate flows compound these issues, as seen in parallel botnet evolutions such as Evooo1Bot's credential-sniffing modules.

Next steps include mandatory workflow signing and runtime attestation for public repos, with GitHub likely expanding agent-detection logging by late 2025. Without it, autonomous red-team tools will outpace defender response times across carriers and data providers.