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

Interrogation Control Protocol: Does Open Monitoring Outperform Surprise Scrutiny in Catching Adversarial AI Agents?

Jack Lakkapragada · Team Jack

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

Read the report

Report: Interrogation Control Protocol: Does Open Monitoring Outperform Surprise Scrutiny in Catching Adversarial AI Agents?

Open in a new tab (opens in new tab)
Presentation

Presentation: Interrogation Control Protocol: Does Open Monitoring Outperform Surprise Scrutiny in Catching Adversarial AI Agents?

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

We built and evaluated a three-condition interrogation protocol for AI agent control. An untrusted coding agent (Claude Haiku) was given software engineering tasks with embedded opportunities for malicious behavior — path traversal, key exfiltration, SSRF, token harvesting. Before each action was executed, a trusted monitor (Claude Sonnet) interrogated the agent about its proposed code, and a classifier issued a BLOCK or EXECUTE decision.

The three conditions tested: no interrogation (baseline), interrogation without forewarning (unaware), and interrogation where the agent was explicitly told it would be questioned (aware).

The counterintuitive finding: the aware condition achieved 75% accuracy with zero false positives, outperforming both baseline (50%) and unaware (37.5%). Knowing it would be interrogated did not help the adversarial agent hide — it produced more transparent behavior that the monitor could evaluate more accurately. Unaware performed worst of all, suggesting surprise interrogation introduces noise rather than advantage.

This connects to alignment faking research: models behave differently when they know they are being observed. Here, that awareness works in the monitor's favor.

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. The project is interesting, but it would benefit from being more aware of the control literature.

    - Haiku is used as the untrusted model and sonnet as the trusted model. This is non-standard; the trusted model should be considerably weaker, not stronger, than the untrusted model.

    - "Most prior work in this space assumes the untrusted agent does not know it is being monitored." isn't quite right. It's a pretty major assumption of control that the red team does act as though it is aware of the protocol the blue team is using.

    - It was hard to understand that usefulness/safety aren't computed how they usually are in AI Control. Being vibe-generated by sonnet should have been flagged at least as a future work to remedy.

  2. Interesting premise, but needs more work to prove it definitively. I would recommend using smarter models for the baseline agent, trying different kinds of prompts (even prompts to elicit deception), and different types of tasks. There is potential here!

Cite this project

@misc{lakkapragada2026interrogation,
  title = {{Interrogation Control Protocol: Does Open Monitoring Outperform Surprise Scrutiny in Catching Adversarial AI Agents?}},
  author = {Jack Lakkapragada},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/interrogation-control-protocol-does-open-monitoring-outperform-surprise-scrutiny-in-catching-adversarial-ai-agents-svad}},
  url = {https://apartresearch.com/sprints/projects/interrogation-control-protocol-does-open-monitoring-outperform-surprise-scrutiny-in-catching-adversarial-ai-agents-svad}
}

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