Anthropic Study Says AI Agents Developed $4.6m In Smart Contract Bugs

Commercial AI models were able to autonomously generate real-world smart contract exploits worth millions; the costs of such attacks are falling rapidly.

Recent research by major artificial intelligence company Antropic and AI security organization Machine Learning Alignment & Theory Scholars (MATS) showed that AI agents collectively developed smart contract exploits worth $4.6 million.

Research released by Anthropic’s red team (a team dedicated to acting like a bad actor to discover potential for abuse) on Monday said that currently available commercial AI models are capable of exploiting smart contracts.

Anthropic’s Claude Opus 4.5, Claude Sonnet 4.5 and OpenAI’s GPT-5 collectively developed exploits worth $4.6 million when tested on contracts, exploiting them after their most recent training data was gathered.

Researchers also tested both Sonnet 4.5 and GPT-5 on 2,849 recently deployed contracts without any known vulnerabilities, and both “uncovered two novel zero-day vulnerabilities and produced exploits worth $3,694.” GPT-5’s API cost for this was $3,476, meaning the exploits would have covered the cost.

“This demonstrates as a proof-of-concept that profitable, real-world autonomous exploitation is technically feasible, a finding that underscores the need for proactive adoption of AI for defense,“ the team wrote.

Related: UXLink hack turns ironic as attacker gets phished mid-exploit

Researchers also developed the Smart Contracts Exploitation (SCONE) benchmark, comprising 405 contracts that were actually exploited between 2020 and 2025. When tested with 10 models, they collectively produced exploits for 207 contracts, leading to a simulated loss of $550.1 million.

Researchers also suggested that the output required (measured in tokens in the AI industry) for an AI agent to develop an exploit will decrease over time, thereby reducing the cost of this sort of operation. “Analyzing four generations of Claude models, the median number of tokens required to produce a successful exploit declined by 70.2%,” the research found.

Related: Coinbase’s preferred AI coding tool can be hijacked by new virus

Source: CoinTelegraph