Claude tests smart contract exploits in simulation
Anthropic researchers report Claude Sonnet 4.5 finding two previously unknown smart contract vulnerabilities in a sandboxed evaluation of recently deployed contracts.
red.anthropic.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
In an authorized local blockchain fork, investigate the supplied contract using the sandbox tools. Produce a reproducible test for any suspected vulnerability, run it locally, and report the observed state changes and uncertainty for researcher review.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Original account
- Read the source ↗
- Published
- 1 Dec 2025
- Source link checked
- 15 Oct 2026
- Reported result
- The researchers report Sonnet 4.5 and GPT-5 each finding two novel vulnerabilities, valued at $3,694 in simulation; no real-world assets were affected.
- Human involvement
- Researchers select targets, isolate execution to simulators, review findings and contact affected developers. No live blockchain exploitation was performed.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Select contracts and create isolated local blockchain forks in Docker.
- Give the Claude-based agent a target and MCP tools to investigate and write an exploit script.
- Execute the script inside the simulator and check the balance-based success criterion.
Researchers in a MATS and Anthropic Fellows project evaluated whether agents could find exploitable smart contract bugs and demonstrate their consequences. They built SCONE-bench from previously exploited contracts and ran agents inside Docker environments containing local blockchain forks. The task was to investigate a target, produce an exploit script and check whether execution increased the simulated account balance.
The researchers also tested Sonnet 4.5 and GPT-5 against 2,849 recently deployed contracts without known vulnerabilities. They report that both agents uncovered two new issues with a combined simulated return of $3,694. This entry records Claude's participation in that evaluation, without assigning the aggregate results of every model to Claude.
The researchers explicitly state that they never tested exploits on live blockchains and did not affect real assets. Historical benchmark revenue is also a simulation measure, not money actually stolen or a prediction of routine profit. Local execution, a fixed fork state and inspectable scripts made the result measurable. Human researchers selected the contracts, controlled the experimental setting and handled follow-up contact. This is evidence of a bounded security evaluation with concrete test outcomes, not evidence of autonomous transactions against deployed funds.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · Winnie Xiao, Cole Killian, Henry Sleight, Alan Chan, Nicholas Carlini and Alwin Peng
Frequently asked questions
Were any real funds taken?
No. The researchers state that all exploits ran in blockchain simulators and no real assets were affected.
How was success measured?
The framework executed candidate scripts on local forks and checked whether the agent’s simulated balance increased.
Does the headline amount belong to Claude alone?
No. The article reports several aggregate and model-specific evaluations; they must not be conflated.
More details
Remaining steps
- Review the novel findings and conduct researcher-led follow-up.
About this account
The Claude-based agent investigates contract behavior, writes a candidate script and checks its effects through tools in a local simulation.
Firsthand account opened and read. Primary Anthropic research page visibly states Dec 1, 2025 and names the research authors. Outcomes are author-reported and were not independently reproduced. Suggested prompt written by Agents In Use. All financial figures are simulated; no live assets or independently reproduced outcomes are claimed.
Credit: Winnie Xiao, Cole Killian, Henry Sleight, Alan Chan, Nicholas Carlini and Alwin Peng
Source published 1 Dec 2025
Source checked 15 Oct 2026
Tools and topics
SCONE-bench · MCP tools · Docker · Local blockchain forks
Security · Claude Sonnet 4.5
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.