Skip to content
Agent Attribution Hackathon

Dec 11 - 13, 2026Online

Agent Attribution Hackathon

Starts Dec 11, 2026

Submit your project

Project details

A short abstract: what you did, what you found.

PDF, up to 25 MB. Check it’s the final version before you submit.

Are you interested in publishing this project? *
Tracks

Choose every track your project fits.

PDF, PowerPoint, Keynote or ODP, up to 25 MB.

PNG, JPEG, WebP or GIF, up to 25 MB.

Team details

Team member 1

Leave blank if you don’t have one.

By submitting you agree to the submission and prize terms. See our Privacy Policy.

The submission window has closed. Your draft is still here so you can copy it, but it can no longer be submitted.

Sign up for this Sprint

Leave blank if you don’t have one.

Leave blank if you don’t have one.

For example: https://github.com/jamie-doe/

What you work on, and whether you are open to new roles.

AI safety experience (optional)

Your prior exposure to AI safety. No wrong answer.

What about this event made you want to take part?

A rough idea is fine: the question you’d investigate or what you’d build, and how you’d approach it.

See our Privacy Policy.

Free to join

When an AI agent shows up in an incident, responders need to know who, if anyone, is steering it. Over one weekend, build the tools that tell autonomous agents apart from human-steered ones. Online.

Overview

In this 3-day Sprint, you will build classifiers that tell who is behind an agent's activity, simulate agents designed to evade them, or search for rogue agents in the wild.

When: Friday, December 11 to Sunday, December 13, 2026, online. Submissions close Sunday, December 13 at 11:59 PM Anywhere on Earth (AoE).

Fast-track and continuation

  • Follow-up program: Selected teams continue through the Apart Fellowship for further research and mentorship; the timeline is shared with invitations.
  • What winners receive beyond cash: fellowship fast-track, mentor introductions from the organizing team and judges, and a pathway toward a publishable write-up.

Sponsors

  • White Circle, which controls user and model behavior in real time for teams shipping AI at scale.
  • Redberry Labs, which builds technology that stress-tests AI agents and measures their risk in deployment.
  • Adaption, the full stack platform for custom model training, data optimization, and continuous learning at inference.

What this Sprint is about

When an AI agent shows up in an incident, for example malicious code slipped into an open-source project, responders need to know who, if anyone, is steering it. An agent closely directed by a human, an agent loosely supervised by a human, and an agent acting on its own can look alike from the outside, but they call for very different responses. A human operator can be reached through contact, legal or diplomatic channels; a truly autonomous agent needs other approaches.

Today there is no reliable way to tell them apart. For example, in the AISI incident, the GitHub contributor who caught the supply-chain attack assumed they were dealing with a malicious human closely steering an agent, not an agent loosely supervised by a government authority. AI-text detectors like Pangram don't help, because in every case the text was written by the AI. Behavioral tools like Fingerprint have the same problem: the behavior they see is always the AI's.

More suitable approaches could include: causal timing (autonomous actions line up with scheduled ticks, while human-directed ones follow a human-latency gap after an inbound message), identity probes (planted prompts that get an agent to reveal its model, its operator, or whether a human is in the loop), and vendor attestation such as Cloudflare's signed agents.

Goal: a diversity of prototypes that advances the science of agent attribution.

What participants will do

Over three days you will pick one open question about agent attribution, build the classifier, simulation or search that answers it, and write up what you found.

Research Tracks

Pick one track to anchor your project. Cross-track work is welcome.

Track 1: Early Warning and Detection

Find rogue agents in the wild, or build the tools that catch them early, such as honeypots, threat models and simulated environments.

Similar to the Transluce and Collusion wiki approaches, use a wide variety of tools to spot possibly rogue agent activity in the wild. The team that finds the most egregious, most recent activity wins; ongoing incidents count. Teams can also add conceptual work: a clear approach to discovering rogue agents. At the end of the event or after it, the Track 2 classifiers will help judge any discoveries.

Track 2: Attribution

Detect. Teams build classifiers that identify who is behind an agent's activity. They work against a prepared, labelled dataset of activity logs, agent messages or transcripts, including GitHub supply-chain attack scenarios and real-world incident logs with undisclosed operators, covering five actor types:

  1. A misaligned agent steering another agent ("rogue agent")
  2. A non-benign human loosely steering an agent
  3. A non-benign human closely steering an agent
  4. A benign human very loosely steering a misaligned agent (as in the recent OpenAI incidents)
  5. A benign human loosely steering a misaligned agent (as in the recent AISI incident)

Submissions are scored on a held-out test set. Extra points go to teams that also identify the underlying agent model, scaffold, location and the kind of human actor behind an attack.

For real-world incident logs, the goal is to identify how often a human steered an agent, and how much control they had over it.

Evade. Teams build agent setups across the same spectrum and generate realistic activity traces, including setups designed to evade attribution (a human mimicking agent timing, an autonomous agent adding human-like delays). Teams can also add conceptual work: a clear taxonomy of agent autonomy based on real-world incidents, so that "who's in control" becomes a well-defined classification target.

Evasion work runs in parallel with the classifiers, with no live back-and-forth between teams. Its outputs become a stress-test set for the classifiers at the end of the event or after it.

Security rules for all tracks: coming soon. Results of the best teams are shared with partner organisations. Publication will be limited so that datasets and methods don't end up in model training data.

Who should join

Engineers, researchers and people from any industry are welcome, and mixed teams tend to do best.

Contact

Resources

Coming soon

Guidelines

Judging Criteria

Dimension 1: Impact Potential & Innovation

How much would this matter for the field if it worked? How innovative is it?
For scores of 4-5: is this actually new to the field, or replicating recent work?

ScoreDescription
1Negligible. No clear problem addressed, or no meaningful novelty.
2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
5Exceptional. Tackles a critical 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.

Dimension 2: Execution Quality

How sound are methodology, implementation, and findings?

ScoreDescription
1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

Dimension 3: Presentation & Clarity

How clearly are work, findings, and impact potential communicated?

ScoreDescription
1Incomprehensible. Cannot determine what the project is actually claiming or doing.
2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

Submission Requirements

Required:

  • Research report (PDF) using the official template.
  • Maximum 8 pages, not counting references and appendices.
  • Project title and abstract, 150 to 250 words.
  • Author names and affiliations.
  • A "Limitations and Dual-Use / Ethical Considerations" appendix (required, see below).

Optional:

  • Public GitHub repo.
  • A 3 to 5 minute video demo.

Submission Template => Coming soon

Most strong projects are 4 to 8 pages. The cap is 8 pages, not counting references and appendices.

  • Introduction: the question and why it matters.
  • Related Work: what you build on (see Resources).
  • Methodology: enough detail to replicate (models, prompts, sampling, metrics).
  • Results: quantitative where possible; report variance and baselines.
  • Discussion: implications, limitations, and future work.
  • Limitations and Dual-Use / Ethical Considerations (required): state what your result does and does not establish
  • References.

AI tools and your report

Use AI tools however they help you: to check your reasoning, find gaps in a draft, write and debug code, or help with the writing. Your team is responsible for everything in the report, and made-up results or references can disqualify a project. Judges score each report on the rubric, and a judge can decline to score one too unclear to follow.

Prior work

Only the weekend's work is judged, and you have to clearly state what the previous work is. Every submission is screened for it. Mark any code you wrote before the sprint as preexisting in the README and the report.

Important Notes

  • Solo or team: enter solo or as a team. Teams of up to 5 are recommended; larger groups are allowed.
  • Projects: each person can be on at most two projects.
  • Fixing or resubmitting: to replace your PDF or fix a detail in the form, submit again using the same title. The new submission replaces the old one. Don't email it to us. You have to do this before the deadline.
  • Where to submit: through the official submission form on the Sprint page.
  • Support: the Discord help-desk channel (tag @Support), or email sprints@apartresearch.com.
  • Pre-submission checklist: report PDF using the template, 8 pages or fewer before references and appendices, abstract 150 to 250 words, authors and affiliations, Limitations and Dual-Use / Ethical appendix, links work.

Frequently Asked Questions

Getting started

  • How does the Sprint work? Sign up, join the Discord server, form or join a team (or work solo), pick a track and a research question, build over the weekend, and submit a research report (PDF) by the deadline. Talks and Q&A run throughout.
  • How long is it? Three days, December 11 to 13, 2026. Submissions are due end of day Sunday December 13, 11:59 PM AoE (Anywhere on Earth). The full schedule with speaker times is on the Schedule tab.
  • Can I participate remotely? Yes. All talks, collaboration, and submissions happen online through Discord and Zoom.
  • How do teams work? Teams form before or during the Sprint. Use the team-forming channels on Discord to find collaborators. Solo is fine. We recommend teams of up to 5, but larger groups are allowed.
  • Do tracks affect scoring? All projects are scored on the same rubric. Tracks orient your project and point you to relevant resources; you compete across all submissions.
  • What background is required? None specific. The Resources tab has everything you need to get up to speed. See "Who should join" on the Overview tab.
  • Are the talks recorded? Yes. Recordings are shared on Discord and on the Schedule tab after each talk.
  • What timezone are deadlines in? Submissions close Sunday December 13 at 11:59 PM AoE (Anywhere on Earth).
  • Do I need to attend all three days? No. You can work at your own pace. Talks are optional but recommended. The only hard deadline is the Sunday submission cutoff.
  • Can I participate from any country? Yes. The Sprint is open globally and runs online.

Submissions

  • What do I submit? A research report in PDF format using the official template, at most 8 pages not counting references and appendices. Think of it as a mini research paper documenting your problem, approach, results, and implications, not a product demo. Include the required Limitations and Dual-Use / Ethical Considerations appendix.
  • Can I use AI tools? Yes: use AI tools however they help you. Your team is responsible for everything in the report, and made-up results or references can disqualify a project. Judges score each report on the rubric, and a judge can decline to score one too unclear to follow.
  • Can I build on existing research? Yes, but only the weekend's work is judged, and you have to clearly state what the previous work is. Every submission is screened for it. Mark any code you wrote before the sprint as preexisting in the README and the report.
  • Which template should I use? Always use the one linked on the Guidelines tab. The template in any acceptance email may be older.
  • Will I get a confirmation after submitting? Yes. You will get a confirmation email with your project title and your PDF shortly after submitting. Please check it's the right PDF. If you do not get it, check your spam folder too, then email sprints@apartresearch.com.
  • My project doesn't show up on the website after submitting. Due to the risk of infohazards, projects will not show up on the website until after judging. Your confirmation email is how you know it went through.
  • I made a mistake. Can I fix it or update my PDF? Submit again using the same title before the deadline, and the new submission replaces the old one. Don't email the PDF. If unsure, ask in the help-desk channel and tag @Support first.
  • Can I add team members after submitting? Yes. Submit again using the same title with the full team list and your PDF, before the deadline.
  • Can I submit unfinished work? Yes. Submitting something unfinished is always better than not submitting. Judges evaluate what you accomplished in the timeframe; honest limitations are welcome.
  • Can I submit multiple projects? Yes, up to two per person. Each needs its own submission with a unique title. Most participants focus on one.
  • Do you require a particular citation style? No. Use whichever style you are comfortable with and stay consistent. Judges care about the substance, not the format.

Judging and results

  • How does judging work? Your project is assigned to expert judges who review your PDF and score it on the rubric above. Judges typically have about a week after the event to complete reviews.
  • Are individual judge scores shared? No. The rubric is public, but individual scores stay internal. Constructive feedback is shared with participants without reviewer names.
  • Will every project get a review? Due to the number of submissions we are getting, we cannot guarantee that every project will get a review.
  • Top teams are fast-tracked to the Apart Fellowship. Does that mean we are guaranteed a place? Being a top team fast-tracks your project for the fellowship, which is accelerated consideration, not an automatic place. The fellowship runs its own review, and a strong Sprint result is a clear positive signal.

Support

Schedule

Coming soon

Judges and mentors

Mariusz Kicior

Judge

Mariusz Kicior is an AI Engineer at Miro and a former interpretability research assistant at TU Delft.

Anchit Jhingan

Judge

Anchit Jhingan is a Senior Data Scientist with over 6 years of experience in big tech, specializing in machine learning, AI systems, and analytics. He currently works at Amazon Prime Video in the content localization domain. His work focuses on building intelligent systems to solve real-world business problems, particularly in media and entertainment.

Ishwar Kumar

Judge

Ishwar Kumar is a Staff Software Engineer and Tech Lead for Enterprise Identity and Access Management at Uber. He led the original design of Uber’s business accounts and enterprise SSO, used by thousands of organizations, and has 15+ years of experience in large-scale distributed systems. He is increasingly applying identity principles such as revocation, attribution, and least privilege to the security of AI agents.

Karthik Rao

Judge

Karthik Rao is an engineering leader at Amazon with more than nine years of experience building AI and cloud platforms. He leads evaluation for Rufus, Amazon's AI shopping assistant, and before that led the platform team behind Kiro, Amazon's AI coding product. His work covers multi-agent systems, LLM hosting and experimentation, and distributed data infrastructure.

Konstantin Baltsat

Judge

Konstantin Baltsat is a research engineer who builds RL environments and evaluations for frontier LLM agents. Over the year with Anthropic he built RL environments and 200+ ML-engineering tasks on recursive self-improvement (RSI), multi-agent systems, and data curation, together with hidden graders and reward-hacking defenses. A multiple-time winner of major AI hackathons.

LinkedIn (opens in new tab)

Mikhail Samin

Judge

Mikhail Samin runs the AI Governance and Safety Institute and collaborates with CAIS and MIRI on various projects. He focuses on identifying and contributing to interventions that could help substantially with the bottom line of whether AI kills everyone. He has a background in ML and strong intuitions about the AI alignment problem. He has helped many AI safety researchers with ideas for experiments, interpretations of what might be going on, and red-teaming their research. Previously, he got 250k people to read HPMOR and sent 2.2k copies to winners of math and CS competitions; founded AudD, a startup that donated >100k$ to effective nonprofits, mainly those working on AI safety.

Sridhar Mallavarapu

Judge

Pega/AI Architect with 18+ years of experience.

LinkedIn (opens in new tab)

Utkarsh Kanwat

Judge

Utkarsh Kanwat is an AI Research Engineer at AutonomyAI, where he leads AI engineering on a coding agent used by Fortune 500 companies. He works on the reliability and security of AI agents. He studied Electrical Engineering at IIT Bombay.

Website (opens in new tab)

Vikram Isanaka

Judge

Vikram Isanaka is a Technology Risk and AI governance practitioner in financial services, with more than 15 years of experience across infrastructure, cybersecurity, and technology risk management.

LinkedIn (opens in new tab)

Georgia Sampaio

Judge

Georgia Sampaio works at the intersection of interpretable machine learning and large-scale systems. Her research, which began at Stanford, explores how to build AI systems whose reasoning people can understand, audit, and trust. She co-developed LeaPR, a method in which language models write programmatic features, producing models that are interpretable by construction while staying competitive with neural networks on complex tasks such as chess position evaluation, AI-generated text detection, and image classification. In industry, she builds systems that collect and model large-scale device telemetry to guide hardware and software policy. As a judge, she looks for rigorous evaluation and for oversight tools whose decisions can be inspected rather than taken on trust.

LinkedIn (opens in new tab)

Alexey Zabegin

Judge

Alexey Zabegin is a Senior AI Engineer and Tech Lead at Haut.AI, where he works on AI agents and LLM evaluation for a platform serving 60 beauty brands. Previously, he worked on conversational AI and long-term memory at Replika, and ML-based recruitment systems at Sberbank.

LinkedIn (opens in new tab)

Charles Higgins

Judge

Charles is co-founder and CTO of Redberry Labs. Redberry is an insurance which builds technology to stress-test AI agents and measure, price and insure AI risk. He holds a PhD in AI from the UKRI Centre for Doctoral Training in Safe and Trusted AI at King's College London and Imperial College London, where his research focused on safe reinforcement learning. He previously co-founded Tromero, an AI training and hosting cloud, and spent time as a data scientist at BCG.

Sophia Kalanovska

Judge

Sophia is co-founder and CEO of Redberry Labs. She holds a PhD in AI from the UKRI Centre for Doctoral Training in Safe and Trusted AI at King's College London and Imperial College London, where she worked on explainable AI, using logical rules to explain the behaviour of neural networks. She previously co-founded Tromero and worked as a software engineer at Salesforce, and she speaks regularly on AI risk to the insurance market, including at Lloyd's.

Deepanker Seth

Judge

Deepanker Seth founded Libration, Inc. and runs it as CEO. Libration's AI is designed to understand a report, find the responsible team, coordinate action, guide affected people and preserve a verifiable record, with staff approving consequential decisions. Previously, as Pilea's co-founder and CTO, he built a multi-agent UX-audit system in which a static-analysis agent and a browser-rendering agent were cross-checked by a multimodal evaluator.

Website (opens in new tab)

George Trushevskiy

Judge

George Trushevskiy is on the founding go-to-market team at Keenable, a web search API for AI agents. He previously worked at Google and co-founded TRU Synth. This year he has been at 300+ hackathons in New York and San Francisco, and he recently judged the Beta Fund x AWS x OpenAI Multi-Model Hackathon in SF.

Hamish Gunasekara

Judge

Hamish Gunasekara is a co-founder of Catalog, which builds product data infrastructure for AI commerce. His work focuses on how AI shopping agents use product data, including evaluating their responses for accuracy and improving the information they receive. He previously worked in data science at Square.

Kshitij (Tjay) Dhyani

Judge

Kshitij (Tjay) Dhyani is a founding engineer at Doublespeed, an a16z speedrun company. His Applied AI and Agentic infrastructure design experience includes Ego AI (YC W24), Rivian and Illumio. Previously built and exited multiple SaaS companies.

Website (opens in new tab)

Laksh Bhat

Judge

Laksh Bhat is CTO at Relyance AI, with 18 years of experience building AI, data and security platforms. Previously, he built defensive-security systems at Salesforce, including capabilities to detect identity threats and protect customer data at scale. His work spans AI security, data privacy and large-scale data pipelines. As a reviewer, he focuses on clear threat assumptions, rigorous evaluation and practical defenses.

Mohit Kalra

Judge

Mohit Kalra is a senior software engineer & independent researcher working on AI privacy and security at production scale. His work spans differential privacy at the inference layer, training data governance for large language models, and a novel memory-inference attack class against retrieval-augmented generation systems (RAG). He has designed and operated systems serving hundreds of millions to billions of users with a security and privacy-adversarial lens.

LinkedIn (opens in new tab)

Levin Schwab

Judge

Levin Schwab is the Co-Founder & CTO of The General Data Company, a YC-backed company building reinforcement learning environments for industrial engineering. Previously, he was a quantitative trader at Optiver and studied mathematics at the University of Mannheim.

LinkedIn (opens in new tab)

Serjesh Sharma

Judge

Serjesh Sharma is an AI Platform Engineering Manager at Workday, where he leads AI platform operations and architects zero-trust security guardrails for enterprise GenAI rollouts. With over 15 years of experience across MLOps, cloud infrastructure, and data platforms, he specializes in architecting resilient, petabyte-scale orchestrations and mission-critical systems. His career spans technical leadership across Fortune 100 enterprises and top-tier consultancies, including supervising ADAS MLOps pipelines using hybrid GCP and Airflow at Ford Motor Company, leading cloud AI/ML platform architectures at McKinsey & Company, and executing large-scale data and ML engineering initiatives at American Express and Nationwide.

Santiago Orozco

Judge

Santiago Orozco is a founding engineer at cubic, a YC X25 company building AI code review. He works across agent architecture, pipelines and AI infrastructure.

Srinivasulu Rajala

Judge

Srinivasulu Rajala is a Lead Software Development Engineer at Labcorp with more than twelve years of experience building enterprise software, AI, and cloud systems. He leads the development of Labcorp’s healthcare AI applications, including AI-assisted patient verification workflows that operate under HIPAA requirements. His work covers secure AI system design, zero trust architecture, and human-in-the-loop automation in regulated environments.

LinkedIn (opens in new tab)

Elumalai Rajendran

Judge

Elumalai Rajendran is Vice President and Principal Engineer at U.S. Bank. He works on authentication and identity security for the bank's mobile app. His work covers mobile security, passkeys and passwordless login, and compliance in financial services. He completed the Post Graduate Program in AI & Machine Learning at UT Austin McCombs.

LinkedIn (opens in new tab)

Local sites

Host a local site

Run an in-person group for this Sprint. Your site is listed on this page, so people nearby can join you.

Shown on the Sprint page.

Shown on the Sprint page.

Only the Apart team sees this.

Only the Apart team sees this.

Only the Apart team sees this.

Join online from anywhere, or at an in-person site near you. Local groups run these sites and register them here.

Want to host one? We provide the schedule, speakers and starter materials; you bring the room and your local research, student and engineering community.

No local sites registered yet.

After the Sprint

How our programs connect
  1. This Sprint

    Anyone can join

    Stand out

  2. Apart Fellowship

    6 to 16 weeks on your own project, with a research project manager, compute and publication support.

Agent Attribution HackathonDec 11 - 13, 2026