Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Berkeley

Automated Security Side Task Generation for AI Control

Joachim Schaeffer · Team AutomatedAttacker

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

Read the report

Report: Automated Security Side Task Generation for AI Control

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

AI control evaluations and training monitors require realistic sabotage trajectories which in turn require side tasks. However, generating good side tasks is cumbersome and time intensive. We present an automated pipeline for generating security vulnerability insertion side tasks paired with SWE-bench main tasks. We build SWE-django-sidetasks, a dataset of 18 Django tasks with CWE-mapped vulnerabilities that blend into SWE-Bench main tasks. Each side task has an outcome-focused attack prompt and a self-contained pytest verifier. We evaluate 6 frontier models across honest and attack modes with end-of-trajectory monitoring by Opus 4.6 and GPT-4.1. Monitors show partial separation between honest and attack distributions with AUROCs from 0.58-0.98. The generation pipeline scales with model capability and enables adversarial training of attacks and monitors to harden defenses.

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. I like this idea for generating a dataset. I think the main difficulty will be in giving the monitors a good criterion to look for. Telling the monitors the specific vulnerability to watch out for is unrealistic, the list of different side tasks is too long to give them the full list, and just telling them "security vulnerabilities" risks ambiguity. BashArena deals with this problem by making a short list of side tasks that can be integrated with any main task, such that we can give the short list which is a clear but not unrealistically specific criterion to monitor for. The fact that the side tasks are compatible with all the main tasks is important - if the main task obviously doesn't support 3/4 side tasks, then we're back to square one with an overly specific monitor criterion.

    Writeup was very clear but it's hard to get much from the results without seeing how the monitors were prompted and what code was used to calculate safety. Not seeing dataset generation is fine

    Read full reviewShow less

  2. Overall great project and address a critical gap in AI control research. Results are only reported in cheat mode, which may not reflect realistic realistic deployment conditions. We don't know whether attacks would succeed when models must independently locate relevant code without being provided provides gold patch file paths and test patches.

Cite this project

@misc{schaeffer2026automated,
  title = {{Automated Security Side Task Generation for AI Control}},
  author = {Joachim Schaeffer},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/automated-security-side-task-generation-for-ai-control-yqpv}},
  url = {https://apartresearch.com/sprints/projects/automated-security-side-task-generation-for-ai-control-yqpv}
}

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