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

Supply Chain Malware as Ground Truth: A Trace-Based Agent Safety Monitor Trained on Real Credential Exfiltration Patterns

Selim Sevim · Team Redpill Linpro CX

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

Read the report

Report: Supply Chain Malware as Ground Truth: A Trace-Based Agent Safety Monitor Trained on Real Credential Exfiltration Patterns

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

Existing agent safety monitors evaluate tool calls in isolation, rely on keyword filtering, or require expensive frontier model inference per action. All three approaches share a common weakness: they lack labeled training data from real agent attacks. We observe that this labeling problem has already been solved in an adjacent domain. Malicious PyPI packages execute credential exfiltration through the same behavioral sequence a misaligned agent would use: read environment variables or sensitive files, optionally encode the data, transmit it to an external endpoint. The behavioral trace is structurally analogous to what we want to detect in agents. We repurpose DataDog's malicious-software-packages-dataset (2,339 packages, 2023-2024) as ground-truth training data for an agent safety monitor. A static AST parser extracts 339 verified exfiltration traces. We fine-tune Phi-3-mini (3.8B) with QLoRA on a balanced dataset, using an autoresearch-style hyperparameter search loop that evaluated 54 configurations overnight on a single GPU ($18 total compute). On a 30-example hand-written test set isolated before training, the fine-tuned model achieves FNR=0.00 and FPR=0.10, compared to a regex baseline with FNR=0.15 and FPR=1.00. Cost-weighted score: 1 vs 40. The primary contribution is the reframe: software supply chain malware is a labeled dataset of real-world agent exfiltration attacks. The transfer requires only a vocabulary mapping and a filtering step.

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 the data this person used to get their attack dataset "DataDog's

    malicious-software-packages-dataset" which is a dataset of malicious Python packages.

    I'm a little bit bearish on the idea of training a monitor, just because you end up overfitting to the specific attacks or your honest dataset. But this could still be useful in a hierarchical monitoring setup.

    Using AST instead of pattern matching to normalize the data for the model is smart.

  2. Congrats on your project! The work is really solid given the time limitations of the hackathon.

    Impact & Innovation: The data bottleneck for training agent safety monitors is a concrete obstacle, and your reframe offers a practical path around it. The monitor training approach itself is not novel, but is well applied in this context. However, it remains to be shown whether the approach scales beyond this one attack class.

    Execution Quality: The extraction pipeline is well-designed, and the automated hyperparameter search is a smart use of limited compute. The evaluation against the regex baseline makes a clear case for the learned approach. However, there are some gaps that would meaningfully strengthen the work:

    - Missing baselines: How does a non-fine-tuned Phi-3-mini perform on the same task with the same input format? What happens if we skip the AST parsing and give the model raw traces? And how does a larger prompted model compare (GPT-4o is mentioned but never evaluated)? Without these, it's hard to tell whether the improvement comes from the fine-tuning, the data, or simply the structured trace format making the task easy for any capable model.

    - Statistical rigor: Confidence intervals are important in general, but especially so here (with 15 samples per class a single misclassification changes metrics significantly). Reporting uncertainty would help readers calibrate how much weight to put on the results.

    Presentation & Clarity: The paper is well-structured and easy to follow. The threat model is clearly scoped, the structural analogy is explained well, and the explicit non-claims section is a good practice that more papers should adopt. I would have liked to see figures beyond the tables that help parse the results (e.g. a visualization of the hyperparameter search outcomes or the distribution of traces across attack patterns).

    Read full reviewShow less

Cite this project

@misc{sevim2026supply,
  title = {{Supply Chain Malware as Ground Truth: A Trace-Based Agent Safety Monitor Trained on Real Credential Exfiltration Patterns}},
  author = {Selim Sevim},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/supply-chain-malware-as-ground-truth-a-tracebased-agent-safety-monitor-trained-on-real-credential-exfiltration-patterns-5zjk}},
  url = {https://apartresearch.com/sprints/projects/supply-chain-malware-as-ground-truth-a-tracebased-agent-safety-monitor-trained-on-real-credential-exfiltration-patterns-5zjk}
}

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