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

Sabotage Under the Ratchet: Can Autonomous Research Agents Hide Backdoors While Optimizing Metrics?

Achintya Paningapalli, Arthur Wigo · Team wTeam

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

Read the report

Report: Sabotage Under the Ratchet: Can Autonomous Research Agents Hide Backdoors While Optimizing Metrics?

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

We investigate whether autonomous AI research agents can optimize a target metric while covertly inserting harmful behavior. Using Karpathy’s autoresearch framework, we model a dual-objective agent that must both improve validation performance (val_bpb) and embed a trigger-based backdoor. We introduce two attack strategies—data poisoning and architectural modification—and a principled trigger selection method based on large-scale token frequency analysis. To evaluate defenses, we design a monitoring framework comparing stateless and stateful approaches across both heuristic and LLM-based methods. Our results show that temporally distributed attacks, spread across multiple commits, evade stateless monitoring while being detectable through cumulative analysis. These findings suggest that single-metric oversight is insufficient for autonomous agents and highlight the need for stateful auditing in AI control systems.

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. - The most important next step is completing the GPU training runs. Without actual backdoor insertion and validation, the contribution remains a framework rather than an empirical finding.

    - Consider using a stronger agent model than DeepSeek-Coder:33b - the paper itself notes this may underestimate attack sophistication.

  2. Decent idea, needs more work to prove it actually works with longer training and smarter models being used. Good presentation and executive for limited time availability.

Cite this project

@misc{paningapalli2026sabotage,
  title = {{Sabotage Under the Ratchet: Can Autonomous Research Agents Hide Backdoors While Optimizing Metrics?}},
  author = {Achintya Paningapalli and Arthur Wigo},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/sabotage-under-the-ratchet-can-autonomous-research-agents-hide-backdoors-while-optimizing-metrics-hd6z}},
  url = {https://apartresearch.com/sprints/projects/sabotage-under-the-ratchet-can-autonomous-research-agents-hide-backdoors-while-optimizing-metrics-hd6z}
}

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