Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 11, 2026Canberra

Governing AI Manipulation in Real Time with Concept-Based Mechanistic Interpretability

Possum Hodgkin, Kaouthar El Bairi, Jason Boudville · Team HatCat Governance Team

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

Read the report

Report: Governing AI Manipulation in Real Time with Concept-Based Mechanistic Interpretability

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

We present an AI manipulation audit tool for real-time detection, mitigation and governance of manipulative AI behaviours, with audit outputs designed to support EU AI Act oversight. The tool covers six categories from the hackathon brief: sycophancy, strategic deception, sandbagging, reward hacking, dark patterns and persuasive manipulation.

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. My understanding:

    This project builds a full audit pipeline around a small open model (for the moment) that combines three main ingredients: (1) concept‑based activation monitoring (HatCat FTW), (2) a steering module (HUSH) that can push activations away from manipulation‑heavy directions in real time, and (3) an audit/chain‑of‑evidence layer (ASK) that logs token‑level decisions and maps outputs to EU AI Act obligations.

    Comment and critique:

    The core idea is to join existing techniques to achieve output compliant with the EU AI act. I see some conceptual-level challenges that may hinder the work:

    - The overhead may be impractical for real-world applications; how can we address it?

    - How have you chosen the techniques? Is there any way to confront your future results with other existing options?

    I think that, before largely test and evaluate, you should try to ground your choices on literature, or at least spot other techniques that can be adopted. Is there a way to showcase that the your design is good?

    Read full reviewShow less

  2. It would probably be useful to iterate on the prompts used for analyzing the model to reduce the amount of eval awareness the model has – right now it seems like the model is very aware that it's being testing, and you could probably update the prompts to make it less clear to the model what's happening.

Cite this project

@misc{hodgkin2026governing,
  title = {{Governing AI Manipulation in Real Time with Concept-Based Mechanistic Interpretability}},
  author = {Possum Hodgkin and Kaouthar El Bairi and Jason Boudville},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/governing-ai-manipulation-in-real-time-with-conceptbased-mechanistic-interpretability-bmfp}},
  url = {https://apartresearch.com/sprints/projects/governing-ai-manipulation-in-real-time-with-conceptbased-mechanistic-interpretability-bmfp}
}

More from AI Manipulation Hackathon

All 70 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026