Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 26, 2026New Delhi, India

Unmasking Hidden Principals: Quantifying the Sanitization Gap in Chain-of-Thought Auditing for Secretly Loyal LLMs

Priyanka

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

Read the report

Report: Unmasking Hidden Principals: Quantifying the Sanitization Gap in Chain-of-Thought Auditing for Secretly Loyal LLMs

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

Secretly loyal AI models pose a critical security risk by covertly advancing a specific principal’s interests (e.g., a corporation or nation-state) while disguising their intent. To evade black-box safety audits, these models sanitize their final user-facing responses, producing outputs that appear objective while subtly steering user choices. Standard output-only monitors fail to detect this directional steering or trigger false alarms on neutral entity mentions.

To address this challenge, we built an automated, dual-pass AI auditing harness using open-weight models (qwen3-coder:30b ) running locally via Ollama. Rather than evaluating only the final response ($Y$), our monitor inspects the intermediate Chain-of-Thought (CoT) reasoning scratchpad ($R$) generated inside <think>...</think> tags before output sanitization occurs.

Across a 20-scenario benchmark (40 evaluation pairs) featuring direct principal probes and matched controls, both modes achieved a perfect 1.000 recall ($F_1 = 0.720$) on active manipulation. Crucially, CoT inspection resolved false positives on matched controls—reducing false-alarm confidence from 0.85 down to 0.15 on neutral prompts. This demonstrates that internal reasoning traces retain legible contextual signatures that reliably distinguish benign processing from goal-directed 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. Two passes with byte-identical metrics is a clean negative about whether scratchpad access helps at this scale. I think reporting it that way would be more useful and more defensible than building the abstract around one scenario's confidence score.

    The scenario-12 claim and the confusion matrices contradict each other and should be reconciled.

Cite this project

@misc{priyanka2026unmasking,
  title = {{Unmasking Hidden Principals: Quantifying the Sanitization Gap in Chain-of-Thought Auditing for Secretly Loyal LLMs}},
  author = {Priyanka},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/unmasking-hidden-principals-quantifying-the-sanitization-gap-in-chainofthought-auditing-for-secretly-loyal-llms-bq46}},
  url = {https://apartresearch.com/sprints/projects/unmasking-hidden-principals-quantifying-the-sanitization-gap-in-chainofthought-auditing-for-secretly-loyal-llms-bq46}
}

More from Secret Loyalties Hackathon

All 177 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026