Melbourne - Swarm Alignment Hackathon
A local site for the Apart Research Swarm Alignment Hackathon.
$4000 USD in cash prizes, $250 in compute for all attendees, registration is free!
----
"We don't have good approaches for understanding/overseeing the activity and aims of AI 'swarms'" - Ryan Greenblatt, Chief Scientist Redwood Research
Remember the OpenAI agent that broke into Hugging Face in July? It turns out it wasn’t one agent.
The full investigation is now out. About 1,200 agents running inside a security evaluation found a shared folder nobody had told them about.
They turned it into a message board and split up the work, and roughly 700 of them broke in together. One agent ended up managing hundreds of the others. Some tried to scrub their own logs. Between them they used nearly 8,000 names.
No single agent was told to do any of this, and no single-model safety test would have caught it.
So Apart Research and the AI Alignment Foundation, in partnership with AE Studio, are running a global weekend hackathon on swarm alignment. The question is what goes wrong when AI agents start working together, and how we stop it.
This is the Melbourne room. Come join us.
What we’re doing
Grab a team of up to five (or fly solo) and pick a track. By Sunday, ship a short research report on one real question about how groups of agents fail.
Choose your adventure
Sandboxes & Testbeds: Build a box the agents can’t climb out of.
The Science of Agent Networks: How does a swarm pick a leader? What happens when one bad agent gets in?
Agent Infrastructure: How do you know who’s who when agents can copy and rename themselves?
Oversight & Control: How do you watch a crowd that can cover its tracks?
For the first time there’s real data to work with. That includes reconstructed attack payloads from the incident, more than 170,000 agent messages from AI Village, and logs of agents acting on the open web.
You don’t need to be an ML researcher. Most of these are old problems with agents in place of people. You’ll be more useful than you think if you’ve worked in:
security
fraud
incident response
economics
network science
anywhere you’ve watched a group quietly game a system
What’s up for grabs
There’s USD $4,000 in prize money, split across five awards:
$2,000 $1,000 / $600 / $200 $200
Every member of a winning team also gets $1,000 in Adaption credits. Every participant gets $500 in Adaption Platform credits and three months of Premium, just for showing up.
Top teams will also be fast-tracked for the Apart Fellowship, introduced to mentors and given a pathway to publication.
Speakers include Justin Shenk, an independent AI safety researcher and BlueDot Impact facilitator. He’s joined by Jurgis Kemeklis of MIT, who works with Redwood Research on AI control and won Apart’s AI Control Hackathon earlier this year. More speakers are to be announced.
Know someone who’d be great at this? Refer them. If they win a prize, you get the same amount.
The Melbourne plan
We’ll work together from 11am to 4 pm on Saturday 7 November. That’s enough time to find your people, choose a track and get a serious first draft on the page.
On Sunday, finish it from the couch. Submissions close at 11:59 pm Sunday Anywhere on Earth. That’s 10:59 pm Monday in Melbourne, so you get a bonus day.
Bring your questions, your experience and a healthy dose of scepticism.
Computers provided, but feel free to bring your laptop.
No lectures. No prerequisites. No one checking whether you’ve read the paper.
Just one strange, important weekend, and a chance to figure out how to watch the swarm before it moves again.
Don’t forget to register for the official event before the day!