Write-up, demo video, and the post Three things that didn't fit in the submission itself.
The demo video — https://youtu.be/jqS_f5T93ak
Four minutes. The console clip is a real 20,000-agent run, unedited and sped up: 20,000 agents, 1,962 distinct thoughts, $1.89 against $19.24 at one call per agent. The run genuinely takes about ninety minutes, because 1,962 real Gemini calls have to happen for the counters to say $1.89.
The write-up — https://medium.com/@uniquedsdave/cache-poisoning-in-collapsed-agent-fleets-b21accf5f0e2
Cache poisoning in collapsed agent fleets. Written for this hackathon, about a vulnerability I haven't seen described anywhere: collapse amplifies prompt injection by exactly the collapse ratio. The saving and the blast radius are the same number. In the largest cohort, that number is 128.
The defence turned out to be structural rather than a filter. A shared answer is addressed by H(kind, role, causal_parents, request), and the request carries only a bucketed projection — eight fields, closed vocabularies, no free text. No attacker-controlled byte participates in a shared address, so cache poisoning here isn't filtered, it's unaddressable.
That implies a constraint that generalises past this project: any design admitting free text into shared reasoning either loses collapse entirely or becomes poisonable. There is no version that keeps both.
The post — https://x.com/i_am_dflame/status/2094628183484280867
Everything in the video regenerates from a command in the README, including the measurements that go against us: collapse costs ~13% of tier-weighted satisfaction, and two baselines with no model at all beat us on that metric. Both are in the repo with the commands that reproduce them.
Log in or sign up for Devpost to join the conversation.