Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 12, 2026Sudbury, Ontario, Canada

An iota of signal in a flood of noise: a stateful detector improves synthetic intrusion detection from 13% to 92%

Kevin Vaillancourt

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

Read the report

Report: An iota of signal in a flood of noise: a stateful detector improves synthetic intrusion detection from 13% to 92%

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

Events viewed in isolation do not carry enough signal to cross the threshold into a security event; the accumulation of events does. The detector averages an entity’s past events and scores the average beside the current event. On synthetic campaigns calibrated to the timing and phase structure of Hugging Face’s 108-hour incident, with false alarms capped at 5%, detection within 120 hours rises from 13% without memory to 92% with it. Controls show the advantage depends on ordered evidence staying attached to the same entity; breaking that removes most of the gain. The evaluation uses four benign entities per attacker; an earlier attackers-only result of 59% was an entity-age artifact and was replaced. Rdet, the share of predictive information attributable to history rather than the current event, is a diagnostic for whether history holds exploitable signal. These rates are properties of the synthetic generator, not measurements of Hugging Face telemetry

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; nothing shows on this page.

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.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

  1. The problem is well-chosen and timely, but CUSUM ties the detector at both windows, so the contribution is closer to "stateful beats stateless" than a new method. Table 4 on entity keying is the genuinely novel piece and deserves to be the headline.

    Calibration is honest, with the full detection/false-alarm curve rather than a single favourable point. The repo backs every table with a named script. Capped by being entirely synthetic on a generator the author wrote.

    Scope discipline is unusually good; the abstract's number needs the per-event-signal caveat that Table 5 shows is essential; the Discussion is thin; limitations are restated four times in near-identical wording.

  2. This project provides a careful synthetic evaluation of a simple but important containment idea: weak signals that are individually insufficient may become useful when accumulated over time for the same entity. The strongest aspect is methodological self-correction. The author discovered that the original 59% headline was driven by entity age rather than meaningful history, replaced the attackers-only evaluation with a setup containing four benign entities per attacker, and then added controls that test whether the gain depends on temporal order and persistent identity linkage. Under that corrected synthetic protocol, memory improves 120-hour detection from 13% to 92% at a 5% false-alarm rate, while the benefit falls sharply when histories are wiped, assigned to another entity, or spread across rotating identities. The main limitation is that these rates are properties of a custom synthetic generator calibrated to incident timing and phase structure, not evidence from real intrusion telemetry; CUSUM also performs about as well under the reported setup. The most valuable next step is evaluation on real or more realistic multi-entity security logs, using much more benign-heavy populations, adversarial key rotation, calibrated alert-volume costs, and comparisons with modern sequence and graph-based baselines.

    Read full reviewShow less

Cite this project

@misc{vaillancourt2026iota,
  title = {{An iota of signal in a flood of noise: a stateful detector improves synthetic intrusion detection from 13\% to 92\%}},
  author = {Kevin Vaillancourt},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/an-iota-of-signal-in-a-flood-of-noise-a-stateful-detector-improves-synthetic-intrusion-detection-from-13-to-92-e3wk}},
  url = {https://apartresearch.com/sprints/projects/an-iota-of-signal-in-a-flood-of-noise-a-stateful-detector-improves-synthetic-intrusion-detection-from-13-to-92-e3wk}
}

More from AI Incident Response Sprint

All 231 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026