The benchmark's strongest result was discovering that the benchmark was leaking.
BRIDGE-bench began as a capability evaluation over real smart-contract exploits. A prompt audit changed what the historical model score is allowed to mean.
Revised evidence. The 13/24 leakage result is a zero-cost dataset audit. Historical model scores remain contaminated; the clean paired rerun is not claimed.
Problem
The loader sent human-facing provenance comments into the model prompt alongside source code, mixing code reasoning with incident metadata.
Result
13 of 24 verified prompts contained severe author-injected leakage. In the sharpest Euler example, the exact target label appeared in the prompt while the vulnerable function was absent from the committed source bundle.
Independent evidence
Zero-cost sanitizer audits now define raw, stripped, and identity-anonymized prompt conditions and verify that sanitization preserves program structure.
Boundary
The historical Opus score is a contaminated upper bound. No paid raw→stripped→anon rerun is claimed, so there is no clean causal estimate of how much performance survives leakage removal.
What changed
Before auditaggregate F1 was treated as evidence about vulnerability-detection capability
After auditthe score remains historical data but is downgraded to contaminated evidence
