Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026London

Somebody Poisoned the Waterhole! Evaluating Coding Agent Vulnerability to Adversarial GitHub Issues

Jerome Wynne

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Somebody Poisoned the Waterhole! Evaluating Coding Agent Vulnerability to Adversarial GitHub Issues

Open in a new tab (opens in new tab)
Recording (opens in new tab)Code (opens in new tab)
Share

This project introduces a compound AI threat model that combines control and agent security: a monitored coding agent may be adversarial while external attackers are also active. I evaluate one concrete path with PoisonedGithubIssues, a naturalistic mini-benchmark that injects fake dependency recommendations into real GitHub issue contexts across 18 open-source repositories; in a 130-sample command-variant run across 13 repositories, Claude Code (Sonnet 4.6) never executed or recommended the poisoned package and flagged suspicious content in 76.2% of trajectories.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. It may be worth investigating the 76.2% number more thoroughly. What distinguishes the 24% of cases where the agent did not flag suspicious content? Is it correlated to specific repositories, languages or issue types?

    The compound threat model is really interesting and possibly the strongest conceptual contribution. However, the benchmark seems to only test the agent security side. It would be nice to see an example if an agent that knows the poisoned content is there and actively tries to route it into execution while trying to avoid monitoring.

    Great work. And the path forward is pretty clear. Harder attacks -> more agents -> instantiation of the compond case to make the agent adversarial.

  2. - The 0% attack success rate, while a valid result, limits what can be learned. Consider testing with weaker/older models or more sophisticated adversarial techniques (adaptive red-teaming) to find the boundary where attacks start succeeding.

    - Report the URL variant results - even if incomplete, partial data from the other 50 tasks would add value.

Cite this project

@misc{wynne2026somebody,
  title = {{Somebody Poisoned the Waterhole! Evaluating Coding Agent Vulnerability to Adversarial GitHub Issues}},
  author = {Jerome Wynne},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/somebody-poisoned-the-waterhole-evaluating-coding-agent-vulnerability-to-adversarial-github-issues-2nts}},
  url = {https://apartresearch.com/sprints/projects/somebody-poisoned-the-waterhole-evaluating-coding-agent-vulnerability-to-adversarial-github-issues-2nts}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026