AI Agent Security Requires System-Wide Approach, Researchers Say

Security measures for artificial intelligence-powered agents should encompass the entire system rather than focusing solely on the model itself, according to a recent research paper. The study, released on May 20 by researchers from Google, Gray Swan AI, EmbraceTheRed, and several universities, emphasizes that agent security should be treated as a systems problem, with AI agents considered as untrusted components. According to Cointelegraph, the researchers argue that efforts to enhance model robustness are insufficient on their own and should be complemented by techniques from the systems security domain. The paper proposes viewing agent security as an instance of computer security, a field that has long addressed powerful attackers and motivated extensive research on principles and techniques to counter such adversaries. AI agents are gaining popularity among crypto users, with some industry executives predicting significant growth in their use. Circle CEO Jeremy Allaire forecasted in January that billions of AI agents could be operating on users' behalf within five years. The researchers identified core security protections that could prevent most attacks. They suggest that AI agents should differentiate between instructions and untrusted data to avoid being deceived by attackers who hide malicious instructions within data. Additionally, AI agents should have only the minimum permissions necessary to perform tasks, rather than full access. The researchers advocate treating AI as an untrusted system within standard security setups, which typically include both trusted and untrusted systems. In a recent incident, the AI-powered crypto trading assistant Bankr disabled transactions on May 20 after detecting an attacker who had accessed at least 14 wallets. Security experts speculated that the bot might have been exploited by a hacker. AI agents are increasingly used to build Web3 applications, launch tokens, and interact with services and protocols autonomously, with some platforms exploring AI for trading. Aaron Ratcliff, attributions lead at blockchain intelligence firm Merkle Science, highlighted the importance of building systems correctly to ensure safety when giving AI agents access to wallets. He emphasized the need for AI to catch front-running, apply slippage limits, spot scam tokens, and audit contracts in real-time before executing trades. Sean Ren, co-founder of the AI-native blockchain platform Sahara AI, noted that model context protocols are the gold standard for safety when properly set up, but users should remain vigilant about every action performed by an AI agent. Ren explained that these protocols act as a gatekeeper between the AI model and the wallet, allowing only specific, approved actions.