AI Safety hackathons 2026
Hackathons on AI safety: alignment, red-teaming, interpretability and model evaluation.
Right now we list 2 open AI Safety hackathons. 2 list a cash prize; the largest is $5,000 (Threat Intelligence & Threat Modelling Hackathon). 2 say solo entrants are welcome. 2 say they are open to entrants in every country. The next submission deadline is 2026-12-07 11:59 UTC (Threat Intelligence & Threat Modelling Hackathon); the latest is 2026-12-14 11:59 UTC (Agent Attribution Hackathon). Updated 2026-11-20.
An AI safety hackathon asks you to study or improve how AI systems behave, rather than to ship a product. Typical projects test a model for harmful or deceptive outputs, measure how well a safeguard holds up, look inside a model to explain a behavior, or build an evaluation that others can rerun.
We list a hackathon here when the organizer's page names AI safety, alignment, red-teaming, interpretability or model evaluation as a theme, track or prize. Many of these events are research sprints that ask for a short written report instead of an app, so check the submission format on each organizer's page. The deadline shown is the submission deadline in UTC.
- Apart Research · Threat Intelligence & Threat Modelling Hackathon · 2026-12-07 11:59 UTC
- Apart Research · Agent Attribution Hackathon · 2026-12-14 11:59 UTC
Questions
Do I need a research background to take part?
Not always. Some sprints say no prior AI safety experience is needed and welcome engineers, students and practitioners. A focused question you can test in a weekend matters more than credentials. Read each event's eligibility notes, since a few target students or specific roles.
What do I submit, code or a report?
It depends on the event. Research sprints usually ask for a short PDF report with your method, results and limitations, sometimes with an optional code repository or video. Other events ask for a working tool or demo. The organizer's page states the format and any page limit.
What makes a good AI safety project?
Pick one narrow claim and test it with evidence. For example, show that a model fails a specific check, measure how often a defense works, or reproduce a published result on a new model. State clearly what you did not test, because judges value honest limits over broad conclusions.
Is red-teaming allowed, and what are the rules?
Many events invite red-teaming, but they usually set rules: use public data, frame the work defensively, and do not publish working attack details that could cause harm. Some ask for a section on dual-use risks. Follow the event's rules and the model provider's terms of use.
How are prizes paid?
Prizes are often paid by the organizer or a partner charity after the event, and payouts can be subject to tax withholding or bank transfer restrictions in some countries. Each listing notes what the organizer says about payment and eligible countries.
Can I enter alone?
Many AI safety sprints accept solo entrants as well as teams. The summary above says how many events in this list welcome solo entrants, and each event's page gives its team size limits.
Other categories
Guides: For students · New to AI · Experienced builders
One email a week
New events that match your topics, every Monday. Unsubscribe anytime.