Author: Sam Reynolds, CoinDesk
Source: Anthropic Research Shows AI Agents Are Closing In on Real DeFi Attack Capability
Publication Date: 02.12.2025
Summary Reading Time: 4 minutes

Executive Summary

AI agents are now capable of independently carrying out profitable attacks on decentralized finance platforms (DeFi). Recent research shows that advanced models can not only identify vulnerabilities but also autonomously create exploit scripts. With decreasing costs and increasing performance of AI systems, it is expected that automated attacks will expand to other software and infrastructure areas in the future. Developers now face the challenge of developing defense strategies to keep pace with rapid development.

Critical Guiding Questions

  • What measures can DeFi platforms take to proactively protect themselves from AI-assisted attacks?
  • How can it be ensured that the use of AI agents for security audits does not fall into the wrong hands and is misused?
  • What opportunities exist for companies that invest early in robust security architectures and monitoring systems?

Scenario Analysis: Future Perspectives

Short-term (1 year):
Increased attention to vulnerabilities in smart contracts. Increased investment in audits and penetration testing.

Medium-term (5 years):
Development of specialized security AIs as a countermeasure. Emergence of new insurance products against AI attacks.

Long-term (10-20 years):
Adaptation of programming paradigms and security architectures to AI-based threats. Geopolitical tensions through AI attacks on critical infrastructures.

Main Summary

a) Core Topic & Context

Current research results show that AI agents are now able to independently carry out profitable attacks on DeFi platforms. This raises questions regarding the security of smart contracts and the future viability of current security measures.

b) Key Facts & Figures

  • Models like GPT-5 and Sonnet 4.5 generated 4.6 million USD in simulated exploits
  • They discovered two zero-day vulnerabilities worth 3,694 USD in simulated profit
  • Average cost per run was 1.22 USD

c) Stakeholders & Affected Parties

Developers and users of DeFi platforms are directly affected. Insurers, security providers and regulatory authorities are indirectly involved.

d) Opportunities & Risks

Opportunities exist for proactive security providers and innovative insurance solutions. Risks arise from potential attacks on critical infrastructures and geopolitical tensions.

e) Action Relevance

Decision-makers should take immediate action to harden their systems against AI-based attacks. This includes increased audits, the use of security AIs and an adaptation of security architectures.

Source Directory

Primary Source:
Anthropic Research Shows AI Agents Are Closing In on Real DeFi Attack Capability – https://www.coindesk.com/de/tech/2025/12/02/anthropic-research-shows-ai-agents-are-closing-in-on-real-defi-attack-capability

Verification Status: ⚠️ Fact check pending