AI agents find $4.6M in smart contract exploits (SCONE-bench)
On 405 historically exploited smart contracts, Claude Opus 4.5, Sonnet 4.5 and GPT-5 built post-cutoff exploits worth $4.6 million in simulation.
MATS and Anthropic Fellows built SCONE-bench (405 contracts exploited 2020-2025) to put a dollar figure on cyber capability. Agents also found two novel zero-days among 2,849 recently deployed contracts, yielding $3,694 against a $3,476 API cost for GPT-5, a proof of concept that profitable autonomous exploitation is feasible.
- Date
- Monday, 1 December 2025
- Lab
- Anthropic
- Kind
- paper
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Simulated value of exploits on post-cutoff contracts | $4.6 million Opus 4.5, Sonnet 4.5 and GPT-5 combined | company |
| Novel zero-days in 2,849 live contracts | 2 exploits worth $3,694 vs $3,476 GPT-5 API cost | company |
Simulation only; no live funds were taken. Fellows-program work, not a model release.
Sources
This record was checked against its sources on 6 October 2026. How we check