This research exposes critical security vulnerabilities in private LLM evaluations conducted by third-party data curators, highlighting significant conflicts of interest and data integrity concerns.

Key Findings:

• Private evaluations lack transparency and accountability mechanisms, creating opportunities for manipulation
• Third-party evaluators face potential conflicts of interest when assessing competing models
• Data contamination and prompt leakage risks are amplified in closed evaluation environments
• Security vulnerabilities threaten the integrity of the entire LLM evaluation ecosystem

Why It Matters: As organizations increasingly rely on third-party evaluations for critical LLM deployment decisions, these hidden security risks could significantly impact product development trajectories and investment strategies across the AI industry.

Peeking Behind Closed Doors: Risks of LLM Evaluation by Private Data Curators