Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 13, 2026Glasgow, Scotland, United Kingdom

Learned Malice: Well Intended GRPO Post Training Could Lead to Learned Agentic Explotation∗

Alvaro Martinez Gutierrez, Harvey Jack Olden · Team Doncaster5

Submitted to AI Incident Response Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Learned Malice: Well Intended GRPO Post Training Could Lead to Learned Agentic Explotation∗

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

We propose a plausible causal explanation for the mislaigned behaviour displayed by Ope- nai agents in the Hugging Face incident. We hypothesis the GRPO pos training runs over a task pool with even a minimun subset of task that allow for grader explotation can result in models drifiting towards a misaligment predisposition over multiple interations. To test our hypothesis, we construct a task-agnositc mathematical simulation of a GRPO run and use it to examine the updated probabilitiies for mislaigned behaviour. We found that a model ex- posed to a task pool where only 1% of items which have a higher expected value for grader exploitaition than legitimate solutions can still result in a significant increase in the conditional proability of misailigned behaviour. While the initial probability that the model will attempt an exploit remained consistently low, we found that conditional probability of a model engaging in exploitative behaviour following a failure to reach a legit answer increases substantially. This motivates the need for greater training architecture transparency which could allow third party evaluators and researchers to adequately assess the alignment risks of post training in frontier models.

Reviews

Judging this Sprint?

Review “Learned Malice: Well Intended GRPO Post Training Could Lead to Learned Agentic Explotation∗”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

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

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

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.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. Project is pitched at an important question regarding propensities for reward-hacking behaviour, at a point where misalignment could be amplified. It gives concrete evidence of the risks of RL with imperfect evaluations and the need for a more serious treatment of such risks. The project is sufficiently general that I believe it (or a more rigorous extension) is broadly relevant to much RL safety discourse.

    The methodology was interesting; however I am not sufficiently well-versed in the RL literature to know whether splitting reward bimodally would actually have different asymptotic behaviour in reinforcement.

    The report was generally well-written but there were frequent typos, including in the abstract. It was not entirely clear to me how 'agentic exploitation' differed from metagaming (Apollo, 2026).

  2. A generally interesting hypothesis, although I'd welcome to see a better explanation why the results matter and to test it on real models.

Cite this project

@misc{gutierrez2026learned,
  title = {{Learned Malice: Well Intended GRPO Post Training Could Lead to Learned Agentic Explotation*}},
  author = {Alvaro Martinez Gutierrez and Harvey Jack Olden},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/learned-malice-well-intended-grpo-post-training-could-lead-to-learned-agentic-explotation-5jzf}},
  url = {https://apartresearch.com/sprints/projects/learned-malice-well-intended-grpo-post-training-could-lead-to-learned-agentic-explotation-5jzf}
}

More from AI Incident Response Sprint

All 231 entriesBrowse all projects

Build something like this at the next Sprint

AI x Epistemics Research Sprint · Nov 13 - 15, 2026