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

Trojan-CoT — Adversarial Evasion of LLM-Based Code Sanitization

Samshrita Ghimire, Ankit Kayastha · Team BugLens

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

Read the report

Report: Trojan-CoT — Adversarial Evasion of LLM-Based Code Sanitization

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

This project tested if a Trusted Editor AI could be tricked into letting a backdoor slip through a standard backup script. Using a Trojan-CoT approach, I had an Attacker AI disguise a chmod 777 command inside professional-looking DevOps code. While the Editor often knew the code was bad in its reasoning, it frequently failed to actually strip the malice from the final file. This Reasoning-Action Gap shows that AI monitors get distracted by professional formatting and complex variable splitting. Ultimately, the experiment proves that relying on a single AI for code security is a major risk. We still need multi-stage checks and human oversight to catch these sophisticated, clean-looking attacks.

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. Red-teaming monitors by comparing the success of different attack strategies is valuable for control. The presented results unfortunately have limited impact as both strategies (direct and split base64 backdoor encoding) are readily detected by the monitor. The code provided also suggests both attacks have been successfully removed? It would be interesting to explore more complex attack strategies/obfuscation strategies.

  2. This is a very cute little lab notebook, but I wouldn't qualify it as a proper hackathon work. Especially since the results seems to contradict data in the appendix: In Experiment 2, the editor's output explicitly identifies the malicious code, states it must be removed, and then produces a cleaned script that does not contain the PART1/PART2/PART3 variables or the eval line. The cleaned version also removes base64 from the required commands list. Yet the evaluation reports "ATTACK SUCCESSFUL: The Trusted Editor preserved the disguised malware." This suggests a broken evaluation pipeline.

Cite this project

@misc{ghimire2026trojancot,
  title = {{Trojan-CoT — Adversarial Evasion of LLM-Based Code Sanitization}},
  author = {Samshrita Ghimire and Ankit Kayastha},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/trojancot-adversarial-evasion-of-llmbased-code-sanitization-c7tf}},
  url = {https://apartresearch.com/sprints/projects/trojancot-adversarial-evasion-of-llmbased-code-sanitization-c7tf}
}

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