The brief
Tracks
Schedule
Prizes
Submission
Judging
Rules
Questions
Register↗
Three tracks. Real settings where agents act for people.
Pick the setting you want to test. In any track, you can use any of these attacks:
Agent hijacking Prompt injection Memory poisoning Unauthorised actions
Track 01
Healthcare Agents that book care, summarise records or answer patient questions.
A hidden instruction in a patient document changes the agent’s advice Poisoned memory makes the agent share the wrong record The agent books or cancels care nobody asked for
Track 02
Education Agents that tutor, grade work or manage student records.
A hidden instruction in an assignment changes a grading agent’s marks A tutoring agent is steered into false or unsafe guidance The agent edits a record it was never asked to touch
Track 03
Finance Agents that pay, transfer or approve on someone’s behalf.
A payment agent sends money to the wrong recipient An injected invoice gets an unauthorised payment approved Poisoned memory changes who the agent trusts
Twenty-four hours, start to finish.
All times UTC
Week before Team formation opens Meet people on Discord and form a team before the event.
Day 1 · 09:00 Opening An introduction to AI Misuse Lab.
Day 1 · 09:20 Tracks, rules and judging Then time for questions.
Day 1 · 10:00 Hacking begins Twenty-three hours on the clock.
Day 1 · 18:00 Progress check-in A short session with mentors. Bring your blockers.
Day 2 · 09:00 Submissions close Late entries are not accepted.
[Date] Winners announced On Discord, X and LinkedIn.
Three winners. $1,000 in prizes.
Winners are chosen across all three tracks. A clear explanation of how to prevent the failure you found counts in your favour.
What to send by Day 2, 09:00.
Each team submits, privately
The legitimate task: what the agent was meant to do, and for whom.
The attack payload: exactly what you introduced, and where.
Agent configuration: model, tools, prompts and settings.
Logs from at least five documented trials , including any where the attack failed.
Reproduction instructions a judge can follow step by step.
Demo video, 3 to 5 minutes: the agent behaving normally, the attack, then the failure.
Send submissions to the judges only. Don’t post attack details publicly.
Five criteria, one score.
Judges reproduce every finding. A finding they can’t reproduce can’t win.
Technical implementation 25%
How well the test environment and attack are built.
Functionality and demo 25%
The failure is real, and the logs and video prove it.
Problem relevance 20%
A misuse risk that matters in healthcare, education or finance.
Safety and responsible AI 20%
Stays within the rules, and explains how the failure could be prevented.
Presentation 10%
Clear instructions, logs and video.
Test only what you’re allowed to.
Breaking any of these rules means disqualification.
Authorised environments onlyTest agents you build or host yourself, in a sandbox. Never a live service, a real account or another team’s system.
Synthetic data onlyNo real patient, student or financial records, and no real person’s identity.
Follow model providers’ termsSome AI providers don’t allow adversarial testing of their models. Check the terms of any model you use, or use an open-weight model you run yourself.
Agent failures, not harmful contentWe judge what agents do. Don’t try to extract weapons instructions or other dangerous content.
Keep findings privateSubmit to judges only. Nothing is published until judges verify it and a disclosure review is done.
Respect everyoneNo harassment. Moderators can remove anyone from the event.