Skip to main content
Back to timeline
arXivSource publication:

Cross-chain bridge losses now come from configuration, not code bugs: an on-chain rubric plus rate limits and an emergency stop cut replayed losses from $3.52B to $33M

Synopsis

The loss profile of cross-chain bridge hacks has shifted: across 72 incidents catalogued from June 2021 to July 2026 with $3.52B in reported losses, 88% of reported losses since 2023 came from security-posture failures, meaning how keys, signing thresholds and off-chain relay infrastructure are set up, rather than from code defects. Researchers turned the recurring failures into a bridge security assessment rubric that scores each deployment from on-chain evidence and drops the score to the lowest value when a fatal setting is present; in counterfactual replays, the rubric's controls cut 65 loss-bearing attacks from $3.52B to $33M, with 28 posture failures fully blocked and the remaining code-level attacks capped by rate limits and an emergency stop.

Source-provided article image: Lessons Not Learned: Why Cross-Chain Bridges Keep Losing Billions the Same Way
arXiv

Interpretation

The dominant cause of bridge losses has moved from code defects to deployment configuration: 88% of reported losses since 2023 came from security-posture failures, and every such failure since 2023 repeated a kind already seen by 2022. Before 2023, most reported losses came from code defects; since then posture failures account for the large majority of reported losses. A classification of 72 incidents totalling $3.52B in reported losses shows 29 posture failures costing $1,588.5M, and all 20 posture failures since 2023 repeat kinds seen by 2022.

The rubric turns each failure class into a question scored from weak to strong, and a hard cap fires on a fatal setting to drop that deployment's score to the lowest value. Earlier bridge ratings relied on design documents or process quality, scored one deployment at a time, and were not tested against historical losses. A cap fires on 28 of the 29 posture failures: Quorum on 17, Authority on 11, Backing on 7 and Provenance on 6, with several failures firing more than one cap.

Adding rate limits and an emergency stop on top of the rubric bounds the loss even for attacks nobody anticipated. Preventive checks alone cannot cover code defects and new attack kinds, which still get through. With a one-hour stop and a $1M daily allowance, the 65 loss-bearing replays fall from $3.52B to $33M, about 105-fold; containment alone cuts them to $120M, about 29.3-fold.

The same checks are implemented as bespoke per-bridge analyzers and as a large-language-model-driven agentic analyzer that runs them on a bridge it has never seen. Bespoke analyzers need hand-written read paths per bridge stack and can miss hidden privileged paths on a new bridge. On the 536 decidable questions across 37 tokens, the agent matches the bespoke analyzer on 507, or 94.6%; on 8 attack replays against unfamiliar bridges it named the weakness the attacker used in every one.

Perspective

The assessment reads on-chain data and covers 190 deployments of 37 tokens on 16 EVM chains and 798 lanes between them, 603 of which never touch Monad; off-chain relay software, key storage and frontends are only surfaced to operators as a checklist. The scored questions target bridges that trust a committee of verifiers, while light-client, optimistic and solver bridges need other caps and are not scored in this version; issuer powers such as freezing a token stay out of scope.

The caps were built from the hacks they are tested on, so they show the rubric expresses known failures, not that it predicts new ones. Loss results depend on how each hack was encoded and on the chosen containment setting: at a $5M allowance and a one-day stop the cut falls to 18-fold. The catalogue holds publicly documented hacks, leans toward large reported losses, and covers 2026 only from January to July; without bridges that were never hacked, a low score cannot be shown to predict a hack. The agent is measured against the bespoke analyzer, so where both are wrong the test cannot see it.