ezo labs research

EZO-R-2026-10 // inference //

How a Chinese-language red-team swarm farms free LLM tokens

Free quota is the target A red-team swarm written in Chinese hunts open model relay panels for keys, default admin passwords and free sign-up quota. Then it tests each key and keeps the ones that work.

What the gateway burst was, by share of requests and share of tokens Horizontal bar chart drawn to scale from zero, two bars for each of eight classes of work in the eleven-minute gateway burst of 6 November, first the share of its 8,648 requests, then the share of its 84.8 million answered tokens. Security-testing swarm 62.3% of requests and 64.3% of tokens; roleplay frontends 13.7% and 7.6%; coding agents 12.2% and 22.4%; content and data pipelines 3.1% and 0.2%; formal-proof pipeline 2.8% and 1.3%; other chat 2.5% and 1.4%; other agent loops 1.9% and 2.7%; key and model probes 1.6% and 0.1%. The roleplay, content and other-chat classes are approximate. Coding agents are 12% of requests and 22% of tokens because each answered request carries a long re-sent history. What the gateway burst was Eight classes of work in the eleven-minute burst on one operator's three accounts. Each class: its share of the 8,648 requests, and its share of the 84.8 million tokens. SHARE OF THE BURST Requests Tokens 0% 20% 40% 60% Security-testing swarm Roleplay frontends* Coding agents Content and data pipelines* Formal-proof pipeline Other chat* Other agent loops Key and model probes 62.3% 64.3% 13.7% 7.6% 12.2% 22.4% 3.1% 0.2% 2.8% 1.3% 2.5% 1.4% 1.9% 2.7% 1.6% 0.1% * Approximate class. The swarm, agent-loop and probe shares are firm.

earlier

  1. How many subagents is too many? We gave ours 12 slots; the median used 1.3.environments
  2. Let your lead agent only plan: fewer turns, same scoreenvironments
  3. Your agent mostly sees one line of each skill. Write that line well.evals

topics

inference
2
environments
3
evals
1

atom feed