Anthropic and Huawei studies reveal critical flaws in Windows vulnerability assessment and LLM memory security

Summary

Two major security studies published last week reveal critical vulnerabilities in both traditional software and AI systems. Anthropic's Mythos model demonstrated that 13 out of 14 Windows bugs classified by Microsoft as "unlikely to be exploited" can actually be weaponized using only publicly available patches, with one exploit achieving full SYSTEM-level control. Separately, Huawei's MPBench benchmark found that over 50% of attacks on LLM agent memory succeed—a single fabricated fact planted in a trusted document becomes permanent "memory" that persists across sessions, with current detection methods catching fewer than half of such attacks.

Source

bskyAnthropic and Huawei studies reveal critical flaws in Windows vulnerability assessment and LLM memory securityundercodetesting.com