Crypto exchanges brace for AI that can exploit software flaws

The crypto industry is moving to shore up defenses after research showed advanced AI agents can autonomously find and exploit software and smart-contract vulnerabilities. Anthropic’s Frontier Red Team experiments prompted renewed focus on how exchanges and custodians manage code risk.

Anthropic and collaborators built a benchmark called SCONE-bench and ran frontier models against 405 historically exploited smart contracts from 2020–2025; the tests produced working exploit paths across a substantial subset of cases and produced a lower-bound simulated haul of about $4.6 million for contracts outside the models’ training cut-off. In prospective scans of recently deployed contracts, the reported average inference cost per contract was roughly $1.22, illustrating that exploit generation is increasingly cost-effective.

For exchanges, the implications are operational rather than hypothetical: AI-augmented development pipelines and third‑party code assistants can unintentionally propagate vulnerable patterns into production systems, increasing custody and transactional risk. Industry reporting indicates major exchanges have been both targets of social-engineering schemes and are reassessing technical controls and access privileges in response to the evolving threat landscape.

In the broader economic and regulatory context, the findings are likely to accelerate adoption of continuous, AI-driven security testing and change the calculus for cyber insurance and compliance. Anthropic and others recommend integrating adversarial AI testing into development lifecycles and publishing defensive benchmarks to help close the audit-to-exploit gap. Regulators and market infrastructures may also demand more demonstrable, continuous assurance from custody providers.

Market analysts expect short-term increases in security spending, faster deployment of AI for defensive tooling, and tighter operational controls at centralized venues. The consensus view is that defenders must adopt the same automated tooling attackers use; otherwise the economics of autonomous exploitation will increasingly favor malicious actors, with broader implications for liquidity, insurance pricing and institutional participation in crypto markets.