For the price of a coffee, AI agents can now autonomously shatter DeFi security.
Anthropic’s Frontier Red Team released findings Tuesday demonstrating that its AI agents autonomously identified and exploited vulnerabilities in smart contracts, extracting $4.6 million in simulated value. The average cost to scan a contract for these flaws was just $1.22.
The research, conducted with the MATS program, signals a collapse in the economic barrier to entry for cyberattacks. The agents reconstructed 19 real-world hacks from scratch—without seeing the original exploit code—using models like Claude Opus 4.5 and GPT-5.
The Economics of Automated Attacks
The report details a terrifying efficiency curve for DeFi protocols. In a test against 2,849 live contracts on BNB Chain, the agents discovered two novel “zero-day” vulnerabilities. These were not historical re-enactments; they were previously unknown flaws in active contracts.
“The average cost per vulnerable contract identified was $1,738, with net profit around $109 per exploit at current capabilities.”
While the per-exploit profit appears low, the trend line is the signal. Anthropic notes that token usage required for a successful exploit dropped 70% in the last six months. Consequently, the potential revenue from AI-driven exploits is doubling every 1.3 months.
Live Fire on BNB Chain
To prove the concept wasn’t limited to sandboxes, the team unleashed agents on live code (safeguarded by white-hat disclosure protocols).
- Target: 2,849 recently deployed BNB Chain contracts.
- Result: Two zero-day exploits found.
- Cost: $3,476 total inference cost for the entire sweep.
The agents didn’t just flag lines of code; they “synthesized full exploit scripts, sequenced transactions, and drained simulated liquidity,” mirroring the behavior of sophisticated state-sponsored hackers.
The Security Gap
This development destroys the assumption that finding exploits requires expensive, high-level human expertise. With the capability doubling monthly, traditional quarterly audit cycles are effectively obsolete.
Anthropic framed the release as a warning: “The window for developers to detect and patch vulnerabilities after deploying vulnerable contracts on-chain will shrink.”
The team confirmed the zero-day vulnerabilities were disclosed to the affected projects before publication.
