Tools
2026-08-18
An AI agent broke into Snowflake's build system in five days. Who wrote the hole is now in dispute.
Security firm Wiz disclosed on 17 August that a pull request merged into a public Snowflake repository on 18 June replaced a safe pattern — passing values through environment variables — with untrusted input dropped straight into a shell command inside a GitHub Actions workflow. Anyone could then run commands on Snowflake's build system simply by opening an issue with a quotation mark in the title. Wiz's automated red-team agent found and exploited it on 23 June and pulled out a Jira API token that gave read access to Snowflake's engineering, security-compliance and bug-bounty projects. Snowflake patched the same day, rotated the token on 24 June, and says no one else got in. (Correction: our first version said GitHub Copilot co-wrote the change and reviewed it as all-clear. GitHub says a human wrote the code and that Copilot Autofix neither reviewed nor contributed to it; the co-author line came from a squash commit that combined several pull requests. Wiz has since softened its own post to say it is unclear whether the change was AI-assisted, and The Register has corrected its headline.)
Why it mattersWhat survived the dispute is still the striking half. An autonomous red-team agent read a public repository, spotted a shell-injection pattern a human had once carefully avoided, wrote a working exploit and walked out with a live credential — five days, no person in the loop. The half that collapsed is a warning of a different kind: 'an AI wrote the bug' is the sentence everyone wants to publish, and it travelled around the world in hours on a co-author line that turned out to prove nothing. Attribution inside a repository is genuinely hard, and a security firm has every incentive to make its finding sound worse than it is. Read the second claim more slowly than the first — including ours.
✓ Verified · 2 sources
▶ Related video: avoiding shell injection in github actions (intermediate) anthony explains #479
Read in the app — free, in 9 languages
Related stories
Apple Music will tell you when a song was made by a machine — but the uploader decides whether to say so
2026-08-21Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship
2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21Stripe has just paid $7.5 billion for a model router. Days later Ramp built one and is giving it away until January.
2026-08-21