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

Towards Early Warning for AI Agent Incidents: Monitoring the Dynamics of Risk Trajectories

Yifan Li, John Gia Bao Luc, Rashik Shahjahan, Kai Ting · Team AI response T1

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

Read the report

Report: Towards Early Warning for AI Agent Incidents: Monitoring the Dynamics of Risk Trajectories

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

Recent AI-agent incidents highlight the need for monitoring systems that can detect emerging unsafe behavior during agent execution, rather than only assess failures after they occur. Motivated by this challenge, we study a modular early-warning framework for LLM agents based on the dynamics of their observable trajectories. The framework first maps a trajectory prefix \(H_t\), containing messages, tool calls, observations, and actions, to a numerical risk-relevant score \(S_t\), and then monitors the evolution of the resulting score process over time. We demonstrate the scoring layer on partial real-agent logs using StepGuard to construct action-unsafe proxy score trajectories, and study the temporal layer on seeded controlled trajectories using pointwise level, recent slope (``momentum''), EWMA, and CUSUM statistics. In the constructed unsafe examples, momentum and CUSUM cross their selected thresholds before the designated unsafe action, whereas level and EWMA cross at that action; neither constructed safe trajectory alarms under the selected cutoffs. These results are illustrative rather than calibrated detector comparisons. Overall, this sprint provides a proof of concept for a score-then-monitor approach to agent safety and motivates systematic evaluation of temporal risk dynamics as a basis for earlier warning and human intervention in deployed agent systems.

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. For a weekend sprint, this study is a clearly scoped proof of concept. The core framing of the study: separating the scorer (by turning a prefix trajectory into a risk score) from the temporal monitor (to watch how that score moves), and treating this score as a time series you can run CUSUM/EWMA/momentum on - is the part I liked the most. The authors are also honest about the study being illustrative, not a calibrated early-warning system. The temporal monitoring code is in the repo and reproduces the table; the StepGuard real-log scoring pipeline isn't committed, so that half is not independently reproducible yet, which is understandable for a weekend and to me a natural thing to include next.

  2. This work does a good job of separating risk scoring from continuous monitoring and shows why looking at the full sequence of agent actions can provide more information than checking individual events. I also liked that the paper is careful about the early-warning results. The main limitation is that the current testing uses constructed scenarios and example thresholds. Testing with larger real-world datasets, realistic false-positive rates, and actual response times would help show whether this approach performs better in practice.

Cite this project

@misc{li2026towards,
  title = {{Towards Early Warning for AI Agent Incidents: Monitoring the Dynamics of Risk Trajectories}},
  author = {Yifan Li and John Gia Bao Luc and Rashik Shahjahan and Kai Ting},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/towards-early-warning-for-ai-agent-incidents-monitoring-the-dynamics-of-risk-trajectories-lxdt}},
  url = {https://apartresearch.com/sprints/projects/towards-early-warning-for-ai-agent-incidents-monitoring-the-dynamics-of-risk-trajectories-lxdt}
}

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