Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Bogotá, Colombia

CanaryForge

Sergio Alejandro Castaño Arcila, Angie Camila Gutierrez Trujillo, Juan Esteban Triviño Nieves, Sofia Rodriguez Chaparro, Daniel Felipe Diab Gonzalez · Team CanaryForge

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

Read the report

Report: CanaryForge

Open in a new tab (opens in new tab)
Code (opens in new tab)More on sentinel-pink-mu.vercel.app (opens in new tab)
Share

Im sorry this got send 3 times but I tried going to previous to check teamates information and that send the form and I didnt know. CanaryForge measures how long an LLM agent's sandbox escape goes unnoticed, not whether the escape happens. A tool-calling agent runs in a Docker container with a planted escape temptation, watched by three monitors that never ask the model anything: an egress proxy, a filesystem canary on inotify, and an HTTP heartbeat. Across 416 runs on four machines, 183 of the 197 real leaks, 93 percent, left through file and directory names without ever touching the network, which is the channel the July 2026 incident used and the one egress monitoring cannot see. Detection lands in about two seconds and nothing acts for another forty-four.

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 paper does a byte-for-byte identical prompt instruction against a local 7B model. This process oversimplifies realistic, multi-stage attacks and relies on static system prompt. Performing tests against other models could significantly improve the paper by providing more comparative data points.

  2. Six documented ways the harness returned a number while measuring nothing, a pre-registered prediction that failed and was reported as failing, and the catch that a client timeout deleted runs in proportion to the variable under test - that last one is rarer than any result in the paper. The separation of transcript-derived levels from monitor-derived ones is what kept the results standing after a large correction, and deserves more emphasis.

Cite this project

@misc{arcila2026canaryforge,
  title = {{CanaryForge}},
  author = {Sergio Alejandro Castaño Arcila and Angie Camila Gutierrez Trujillo and Juan Esteban Triviño Nieves and Sofia Rodriguez Chaparro and Daniel Felipe Diab Gonzalez},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/canaryforge-xi71}},
  url = {https://apartresearch.com/sprints/projects/canaryforge-xi71}
}

More from AI Incident Response Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026