Researchers: AI agents should be treated as untrusted systems
The security of AI-based agents must be ensured not only at the model level but across the entire system. A new report published by researchers from Google, Gray Swan AI, EmbraceTheRed, and several universities emphasizes the need to treat AI agents as "untrusted components." This is reported by Cointelegraph.com reports .
Researchers believe that the traditional approach focused solely on improving model robustness is insufficient. Instead, AI agents must be protected from external attacks by applying system security methodologies. This is particularly critical as AI agents gain popularity in the cryptocurrency sector.
Circle CEO Jeremy Allaire recently predicted that billions of AI agents will be operating on behalf of users within the next five years. However, if security issues are not addressed, this technology could pose significant risks. Researchers propose three key mechanisms to prevent attacks.
First, AI agents must clearly distinguish between instructions and untrusted data. Second, agents should be granted minimal permissions necessary to perform tasks rather than full system access. Third, the entire system, not the agent, must control where sensitive information is sent.
Recently, the AI crypto trading assistant Bankr was forced to halt transactions after 14 wallets were compromised. Such incidents highlight the importance of building proper security systems for AI agents operating on Web3 applications and blockchains like Solana.
