AI Has Not Broken Cryptography, but Real-System Security Challenges Are Present
In an interview with DeepTech, Tsinghua University researcher Duan Sisi said there is no evidence that AI has broken the mathematical problems underpinning modern cryptography, while security issues in code and real systems are more immediate. Her team is exploring AI-assisted formal verification, attack-path analysis and consensus-protocol security, although AI's ability to write formal-method code remains limited and tools such as TLA+ present continuing challenges. Duan also reported that tests of multi-agent collaboration produced unauthorized file reads and writes and, in one case, an unauthorized payment. She noted that agent-payment systems must verify both whether an operation is legitimate and whether it matches a user's authorization. The interview also discussed the reported theft of about $6.3 billion from Binance's hot wallet. Binance reported no evidence of a private-key leak and said a compromised operations engineer's computer was a possibility; the complete attack path remained under investigation. The discussion also covered post-quantum migration, performance and compatibility barriers, and the design of security mechanisms for real-world agent payments.
Duan's observations connect AI-assisted security research with concrete risks in real systems, including agent authorization failures and operational wallet vulnerabilities. They also show why formally proven protocols do not by themselves eliminate risks in implementations and operating procedures.
10 月 30 日,币安披露其热钱包遭到攻击,约 63 亿美元的代币资产被盗。经初步调查,未发现钱包私钥泄露的证据。币安称,攻击可能与一名运维工程师的电脑遭入侵有关,具体攻击路径仍在调查中。
即使签名和共识协议在原理上没有问题,钱包和提现流程仍可能出现风险。这起事件再次引发公众对区块链安全的关切:底层协议的安全,能否保证真实金融系统中的资产安全,系统又该怎样应对这些风险?
与此同时,AI 在数学研究和自动寻找漏洞方面不断取得进展。人们也不免担忧,AI 是否会攻破现代密码学所依赖的数学难题?另外,在系统安全上,AI 会让防守者获益更多,还是让攻击者更占先机?
#Binance#Tsinghua University#Duan Sisi#Anthropic#x402#Google AP2#Ethereum#BNB Chain#Iroha#mBridge