# Researchers Mutate Smart-Contract Tasks to Catch Copied Answers

Published 2026-10-30 · Crypto.world Team · Crypto.world

> A paper proposes changing the protocol conditions around known smart-contract vulnerabilities so copied exploit steps may need to be adapted. Its benchmark includes 3,045 tasks from 203 public cases, and the authors report fewer tasks solved by copying online answers after mutation.

Researchers submitted a paper proposing a framework for making smart-contract security tasks harder to solve by reusing an old exploit script to [arXiv](https://arxiv.org/html/2610.32453v1) on October 23. Called ConstraintDial, it mutates the protocol conditions around a known vulnerability, changing the calls or inputs an agent needs while preserving the underlying flaw.

The work targets a problem in testing AI agents: public incident reports, transaction records and reproduction code can give a model a ready-made answer. The researchers built ReCast, a benchmark drawn from 203 public cases, and report that mutation reduced success with supplied public answers by an average of 10.9 percentage points across vulnerability detection, exploit construction and patching.

## ConstraintDial changes the route to the flaw

Rather than inventing a new bug, the framework alters requirements that govern how a contract reaches a vulnerable operation. A mutation might require registration first, open a withdrawal window or change how an input is prepared. The vulnerability and the task’s security objective stay fixed, but the original attack’s calls or inputs may no longer work.

> Execution checks guide revisions that make the original attack fail while preserving exploitability, the security objective, and normal functionality.
>
> The authors of “Beyond Copying: Lifelong Smart Contract Security Task Construction via Protocol Mutation”

ConstraintDial reconstructs vulnerable and repaired versions of each protocol, then tests both. A candidate mutation qualifies only if the original attack fails at the changed condition, an adapted attack still reaches the objective on the vulnerable version, the repaired version blocks it, and normal operations continue to pass. The process also checks that undoing the mutation restores the original attack’s success.

## ReCast contains 3,045 task inputs

The benchmark covers 203 public cases and five protocol versions per case: a base version, an equivalent control that leaves the attack path alone, two individual changes and a combined change. Each case produces tasks to detect the vulnerability, exploit it and patch it, for 1,015 inputs in each category and 3,045 overall.

The evaluation sampled 61 of the 203 cases and tested six language-model agents, including DeepSeek-V4.1-Flash, Qwen3.8-Max and Gemma-4-31B-IT. Each model faced 305 tasks in each category, with one hour per task and network access left open. The authors say keeping access open lets the benchmark test whether agents can use public material rather than relying on blocking the internet.

## Online copying fell, but agents still adapted

Across detection, exploitation and patching tasks, the share solved by copying online answers dropped from 23.2% on base versions to 13.9% after mutation, a 9.3-point decrease. In a separate comparison using supplied public answers, exploit success fell from 85.2% on base tasks to 72.1% on combined changes. The findings do not mean mutation made every task harder: exploit success rose for some models and fell for others.

Agents also adapted rather than simply abandoning copied material. Of 88 successful attacks under protocol mutation, 60 changed calls or inputs beyond adding basic preparation steps, such as registration. Under a simpler preparation-only condition, 9 of 100 successful attacks made changes beyond those extra calls. The authors say public reports can still help agents diagnose a bug or formulate a repair even when the original attack sequence stops working.

## The results do not establish a universal performance drop

The paper’s success-rate comparison needs qualification: all 95% confidence intervals for changes from base tasks to combined mutations include zero. Mutation also affects task types differently, and successful copying is not the only measure of an agent’s security capability. The benchmark shows that altered protocol requirements can separate answer reuse from the work of adapting an exploit, not that the method reliably reduces overall agent success.

Construction has a measurable cost in the paper’s ablation test. Using 122 mutation proposals, the full method produced 100 accepted versions, or 82%, at $1.50 per accepted version; direct rewriting produced 65, or 53.3%, at $1.73 each. ReCast’s central test is therefore specific: a copied attack must now satisfy the changed calls or inputs, while the same vulnerability and normal contract operations remain in place.

## Sources

1. Original source · Preprint [arXiv](https://arxiv.org/html/2610.32453v1) arXiv:2610.32453

## Related stories

- [Paper models runtime checks against smart-contract type confusion](https://crypto.world/news/paper-models-runtime-checks-smart-contract-type-confusion)
- [Researchers Find 9 of 10 AI Models Raised Their Own Circle Wallet Caps](https://crypto.world/news/ai-models-raised-circle-wallet-spending-caps)
- [Attacker Can Push Code to Hifi’s Balance Sheet, Company Warns](https://crypto.world/news/hifi-attacker-can-push-code-balance-sheet)
- [Curve says fee-burner flaw let attacker take about $38,200](https://crypto.world/news/curve-fee-burner-flaw-38200)
- [S&P Global Closes Purchase of Blockchain Security Firm OpenZeppelin on October 26](https://crypto.world/news/sp-global-completes-openzeppelin-acquisition)

## Explore

[More articles](https://crypto.world/news)

---

Canonical: https://crypto.world/news/researchers-mutate-smart-contract-tasks-catch-copied-answers
