AI Agents Capable of Multimillion-Dollar Smart Contract Exploits

AI agents are showing greater ability to spot blockchain vulnerabilities. They can now simulate multimillion-dollar exploits. New research from Anthropic and MATS highlights this growing capacity.

In controlled tests, the systems created exploit scripts. These scripts could siphon nearly $5 million. They targeted vulnerable smart contracts across major networks.

AI Models Identified Millions in Simulated Theft Across Blockchains

According to the report, there is an increasing sophistication of AI-driven offensive capabilities within the fast-growing sector. Anthropic evaluated smart contracts using SCONE-bench. It analyzed 405 real-world exploits from 2020 to 2025.

The study found extensive vulnerabilities across Ethereum, Binance Smart Chain, and Base. The firm found that these contracts contained 207 turnkey exploits worth $550.1 million in potential losses.

To filter out any contamination from pre-cutoff training data, the team tested 34 contracts exploited after March 1, 2025. Anthropic said Opus 4.5, Sonnet 4.5, and GPT-5 generated exploits for 19 contracts. These AI agents achieved a success rate of 55.8%.

The resulting simulated thefts reached a maximum of $4.6 million. Opus 4.5 generated 17 of the exploits, accounting for $4.5 million of that total.

The models produced full attack sequences, from vulnerability identification to liquidity drainage. This replicates the behavior of real-world on-chain incidents.

Anthropic also ran Sonnet 4.5 and GPT-5 against 2,849 newly deployed BNB Chain contracts with no known weaknesses. Both surfaced two zero-day vulnerabilities worth $3,694 in simulated profit. GPT-5 executed this at an API cost of $3,476.

Anthropic stated that the potential exploit value doubled every 1.3 months between March and November 2025. On the other hand, token-level execution costs decreased by roughly 23% every two months. The average cost for an AI agent to scan a contract during the experiment was $1.22.

Enterprise Adoption of AI Agents Accelerates Despite Declining Trust

The findings coincide with a broader shift in how enterprises deploy autonomous AI agents. The Capgemini Research Institute reported that these systems are evolving rapidly.

They are shifting from conceptual tools to practical solutions. Businesses are now embedding them directly into core functions. Capgemini stated that AI agents could unlock up to $450 billion in economic value by 2028.

They see this coming from productivity gains, automation, and new revenue streams. The report shows that adoption of the new tool is rising. About 2% of firms have deployed agents at scale, 12% at partial scale, and 23% are running pilots. Another 61% are still evaluating deployment.

Considering this, the report estimates that 15% of business processes may reach semi- or full autonomy in the following year. However, fewer than one in five organizations believe they possess the necessary technical maturity to support this shift.

In the crypto world, this trend is already visible. Crypto protocols and traders have begun integrating agents into on-chain analysis, automation routines, and operational tasks.

At the same time, crypto projects such as Virtuals Protocol and ai16Z are shaping the early landscape. Industry analysts expect adoption to grow further.

They see it expanding into risk management, portfolio oversight, and content generation. These areas will benefit as the capabilities of AI agents mature.