Study tests whether LLMs can diagnose Solana bridge flaws, not just flag them
A new arXiv paper reports that open-weight models often missed exploitable bugs in a Solana cross-chain bridge, with results varying sharply by input type.
In plain words
Cross-chain bridges move tokens between blockchains and have lost large sums to bugs. The authors built a test bridge with known flaws and asked large language models to find them. The models did best when code comments described the flaw and worst when judging raw transactions.
What was announced
Researchers behind the arXiv preprint implemented a native Rust lock-and-mint bridge on a local Solana validator, creating four exploitable variants: two modeled on the Cashio and Wormhole incidents and two representing vulnerability classes documented in Solana security audits. Paired vulnerable and secure programs, successful attacks, benign sessions, and blocked attempts let them separate correct diagnoses from unsupported alarms. They ran six configurations of five open-weight models, collecting 16,800 responses, and had two blinded LLM graders score 7,200 answers on vulnerable inputs.
Results differed sharply by condition. Without comments, models diagnosed 24% of vulnerable source inputs and often flagged the secure bridge as well. Transaction-level diagnosis was high for the missing-signer scenario but averaged 6% across the other three. When original comments, some of which describe the flaw, were included, source diagnosis rose to 60%. The authors say the findings underline the importance of distinguishing genuine vulnerability diagnoses from unsupported alarms in bridge security review.
What it does not show
- The evidence is the paper's abstract only; full methodology, model identities, and peer-review status are not available in the supplied passages.
- The test covers one bridge implementation with four flaw variants on a local validator, so results may not generalize to other bridge designs or production deployments.
- Diagnoses were graded by LLM graders, and the abstract does not report independent human verification of the grades.
For security and IT teams
- Scope
- The study concerns a local Solana validator test environment and open-weight models, not a deployed bridge or a specific commercial product.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.