Deep Tide TechFlow news: On July 20, US White House AI adviser David Sacks said Kimi K3 has fixed 15 critical security vulnerabilities, while Codex and Fable refused to process related tasks due to “network guardrails” restrictions. He believes that if Chinese models are able to carry out such cybersecurity tasks, then if US models continue to face even stricter limits, their competitiveness may be weakened.

Sacks also cited a Hugging Face example, saying that US frontier models, when analyzing an AI-driven network attack, could not handle requests that included real exploit payloads due to security restrictions. The team ultimately switched to running GLM 5.2 locally, believing that overly strict guardrails may actually undermine defensive security capabilities.