Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Albarrobo, Chile

Blue Beam

Patrick Dugan · Team MoralityLab

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

Read the report

Report: Blue Beam

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

Presentation: Blue Beam

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

BlueBeam is a control-protocol research stack for risky software-agent behavior under imperfect monitorability. We built a benchmark-safe pipeline around BashArena-style trajectories that combines cheap risk heuristics, latent-sidechannel analysis, corruption-gap calibration, token-fed SAEs, and cross-referenced dictionary probes. The strongest current signal is orchestration-linked multi_step_risk_escalation, with useful supporting structure from manipulation-family probes such as unauthorized_config_edit and secret_touch_drift.

The positive result is that latent and probe-based analysis can rank behavior families and layers that separate benign from compromised conditions without saturating into hard-block behavior on benign traffic. The negative result is also included: a family-specific TinyLoRA follow-up did not yet translate into a strong controller-side intervention gain, so we present that as future work rather than overclaiming. We packaged figures, methodology, SAE outputs, and future- work notes in the submission bundle, and published a hosted artifact dashboard here: https://bluebeam-app.vercel.app

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 calibrated controller idea is the strongest thread; build the paper around that.

    - Focus the paper on one clear contribution with strong evidence rather than documenting an entire exploratory pipeline.

  2. I think the main thing I would like is more numbers/metrics -- often I found myself wondering whether results/data were noise or actual signal. I think this was a rather ambitious project.

    The other issue is that you're the one who tampers with the model in a specific way and then builds a detection system to find tampering and kind of that same axis. So it's unclear how well this would generalize and it does not seem to me as if it would generalize that well. You didn't talk much about generally baselining the original model.

    I'd be interested in seeing some adversarial evaluations as well. It would be interesting to see a harmless fine tune and see if was not flagged.

Cite this project

@misc{dugan2026blue,
  title = {{Blue Beam}},
  author = {Patrick Dugan},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/blue-beam-yf5l}},
  url = {https://apartresearch.com/sprints/projects/blue-beam-yf5l}
}

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