Anthropic AI Agents Replicate $4.6M in Historical Smart Contract Exploits

The statement is assessed as 'likely_true' with high confidence based on overwhelming and consistent evidence from multiple high-authority sources, including primary reports from Anthropic itself.1. **Primary Source Corroboration:** The most credible evidence comes directly from Anthropic, the organization that conducted the research. The summary of a primary source from Anthropic's red team explicitly confirms that its "AI agents produced exploits for historical smart contract vulnerabilities, with the value of the 'simulated stolen funds' reaching a maximum of $4.6 million." Another primary source, described as the original research report, also states the AI agents identified exploits valued at $4.6 million. These sources directly support every component of the claim.2. **Consistency Across Sources:** The findings are consistently reported across multiple secondary sources with high relevance, including news reports from Cointelegraph, Incrypted, and The Register. All relevant sources cite the same $4.6 million figure and attribute the research to Anthropic's AI agents.3. **No Contradictory Evidence:** There are no contradictions among the relevant sources provided. The only source that does not support the claim was correctly identified as irrelevant to the specific topic of smart contract exploits.4. **Specifics of the Claim:** The statement's use of the phrase "replicated historical smart contract exploits" is well-supported by the primary source mentioning that the AI "produced exploits for historical smart contract vulnerabilities." The value of "$4.6 million" is also consistently affirmed.In summary, the claim is directly substantiated by the researchers who performed the work and is widely corroborated by independent reporting. The high degree of authority and consistency among the sources results in a very high probability that the statement is true.