Playbooks · Coding tasks
Coding-agent traces in advisory fixes
Written November 10, 2026
A preprint mining GitHub Security Advisories found traces of coding agents in the fixes for hundreds of advisories. It counts traces, not quality — and association is not proof that agents fix bugs better.
The problem
A team deciding whether to let agents touch security fixes has mostly marketing to go on. Vendors say agents fix vulnerabilities; skeptics say agents introduce them. Neither side has had much evidence of what agents actually do in real fixes, in public repositories, where the work can be checked.
What the study found
The paper, Agents on the Patch Path, submitted November 6, looks for observable evidence of coding-agent involvement in commits referenced by reviewed GitHub Security Advisories. The authors built a mining pipeline that flags agent evidence through known agent accounts, explicit commit-message attribution, repository instruction files, and agent configuration artifacts. Linking those findings to advisory-referenced commits gives 469 valid candidate advisories across 189 repositories, including 448 with direct attribution and 21 supported only by repository-level evidence. Claude is the most frequently observed tool family, followed by Copilot and Cursor.
The candidate advisories include high- and critical-severity vulnerabilities, so this is not a corner of trivial fixes. The authors also compared review timing: against advisories published in the same months, candidate advisories show a modestly longer mean review lag, mostly from timing within the same disclosure routes rather than from a different mix of routes. Against a historical baseline, they show substantially shorter mean review lags, with both route mix and within-route timing contributing.
What it does not show
- Association is not causality. The study finds agent traces in fixes. It does not show that agents caused the fixes, that they fixed them faster, or that they fixed them better. A repository that uses agents heavily will leave agent traces in its fixes whatever the agent contributed.
- It does not measure fix quality. The paper counts involvement and compares review timing. It does not grade whether the resulting patches were correct, complete, or free of new problems.
- Attribution is the authors' method. Direct attribution means a commit names an agent or comes from a known agent account; repo-level evidence means the repository shows agent adoption without a direct link to the fix. Both are signals, not proof of who — or what — wrote the code.
- It is a preprint. The paper has not been peer-reviewed, and its numbers are the authors' own.
What a team can take from it
The useful finding is a modest one: coding-agent involvement in security remediation is already observable across hundreds of advisories and repositories, including severe ones. That means agent-written code in security fixes is no longer a hypothetical to plan for — it is already in the patch history of real projects. The questions the study leaves open are the ones a team still has to answer for itself: who reviews agent-written security patches, how thoroughly, and what happens when the agent's fix is wrong.
Limits
- The paper is a preprint, not peer-reviewed work.
- The study counts traces of agent involvement in fixes. It does not measure whether agents make fixes better, faster, or safer.
- The review-timing comparison cuts two ways — longer than same-month advisories, shorter than the historical baseline — and the paper attributes the difference to timing and route mix, not to the agents themselves.
Related playbooks
- Grok Build: SpaceX's $8-a-month coding agent
- MCP Inspector 2.15.0 fixes four vulnerabilities
- Pick a memory server for your coding agent
- Cut coding-agent costs with TokenEconomizer
Sources
Back to: Coding playbooks hub · Also: Secret scanning now catches Actions log leaks · Apiiro: 1,122 Claude settings files leak secrets