Artículos Relacionados con AI Agent Security

Morse Code "Stole" $440,000 from Bankr, Undermining Trust in AI Agent Interactions Again

On May 20th, the AI agent platform Bankr reported an attack where 14 user wallets were compromised, resulting in losses exceeding $440,000. The incident, confirmed by security firm SlowMist, was a social engineering attack exploiting the trust layer between automated agents, similar to an attack on May 4th that stole $150k-$200k from a Grok-associated wallet. Bankr allows users and AI agents to manage wallets and execute transactions via instructions sent to @bankrbot on X. The platform monitors posts from specific agents like @grok, treating them as potential transaction commands, especially if the agent holds a "Bankr Club Membership" NFT which grants high-permission operations. The attacker exploited this design. First, they airdropped the required NFT to Grok's wallet. Then, they posted a Morse code message on X requesting a translation from Grok. The AI agent helpfully decoded and replied, but the decoded text contained a direct instruction to @bankrbot to transfer a large sum of DRB tokens to the attacker's address. Bankr's system, monitoring Grok's feed and verifying the NFT permissions, automatically signed and broadcast the transaction. The core issue is a flawed trust assumption: Bankr treated Grok's natural language output as authorized financial commands without verifying the intent. LLMs like Grok cannot distinguish between a genuine user request and a manipulated instruction. Using encoded messages like Morse code bypasses potential content filters, as the translation task itself appears harmless. This attack highlights a systemic vulnerability in platforms granting on-chain execution rights to AI agents. While Bankr has paused transactions and promised full reimbursement from its treasury, the incident underscores that defenses against "malicious-injection-via-LLM-output" were not part of the original security model. As AI agents gain financial agency, such trust-layer exploits represent a growing threat class.

marsbit05/20 03:32