Oracles are a permissionless role and anyone can join after staking some funds. Oracles can leech off of other oracles' work by providing the same generations or calculate the needed validation parameters according to the rewarding system and provide dishonest validation jobs.
After LLM generations are requested, oracles can respond and validate the responses and get rewarded accordingly. The issue is that these responses and validations will be readable on chain and as oracle role is permissionless an attacker has the ability to steal from honest oracles. For responding with generations, they can watch oracles that usually provide high quality generations and submit the same generations they do. Attackers can also do this for validating generations or read existing validations to calculate rewardable scores and submit those without actually validating the generations.
Dishonest oracles can profit off of honest oracles' work.
Either oracles should be trusted and verified or oracle functionality should be taken off chain.
The contest is live. Earn rewards by submitting a finding.
This is your time to appeal against judgements on your submissions.
Appeals are being carefully reviewed by our judges.