The Security Roundup·Fri, 3 Jul

Tokenization flaws and AI-code risks dominate

Today's cybersecurity news is split between structural vulnerabilities in LLMs and the real-world consequences of AI-generated code. A new paper shows BPE tokenization creates exploitable gaps in safety alignment, while another researcher found a guardrail bypass in ChatGPT's file download flow. Meanwhile, Dawnguard's funding round highlights the 13x surge in AI-code vulnerabilities, and the US government confirms another breach.