Binance's $6.34 billion theft case: How did the hacker use internal AI agents to infiltrate?

Binance disclosed that a $6.34 billion crypto theft, the largest in its history, was carried out by hackers who exploited the exchange's internal enterprise AI agent rather than traditional malicious scripts. After breaching an engineer's workstation through a fake job-interview lure, attackers inherited the engineer's access to an internal AI assistant that held broad cross-service read permissions, letting them map internal topology and locate the wallet-signing service. Rather than stealing private keys from HSMs, the hackers used the AI to obtain high-level service credentials and forced Binance's automated signing engine to authorize large transfers, which were batched on-chain via Ethereum's EIP-7702 account-abstraction standard and initially mistaken for routine wallet rebalancing. Mace AI, which analyzed the attack, said the case exposes a "privilege mismatch" in which AI agents hold far more system access than the employees who use them, and warned that most firms are unprepared for this new attack surface. The incident matters because it shows prompt-level manipulation of trusted internal agents can bypass perimeter and endpoint defenses, and the focus now shifts to whether enterprises enforce least-privilege rules and auditable guardrails on autonomous agents.