Tag
The article highlights a safety flaw in the cross-thread memory implementations of ChatGPT, Claude, and Gemini, where user-provided information from one thread may not be immediately available in another thread, potentially leading to dangerous advice, especially for unaware users.
This blog post explains the current double free detection mechanism in glibc malloc's per-thread tcache, identifies a flaw that allows cross-thread double frees, and proposes a potential fix using randomized keys per thread.