White House AI adviser Sachs criticizes US model reviews: Kimi K3 fixes 15 security vulnerabilities, but US models’ self-limitations are too strict

White House AI adviser David Sacks pointed out that Kimi K3 fixed 15 critical security vulnerabilities within days, while the US-based models Codex and Fable missed handling them due to “cybersecurity guardrails.”
(Background: Before topping the front-end code leaderboard, Kimi K3 beat Claude Fable 5 in real-person blind tests)
(Additional context: Did Kimi K3 push US AI giants to the edge? Experts predict Anthropic is rushing Opus 5, while GPT-6 may be released early)

Table of contents

Toggle

• What are the 15 critical vulnerabilities?
• The “guardrail effect” of Codex and Fable
• The AI speed race in the same week

In a single post, White House adviser David Sacks laid bare the “self-imposed limits” in the US AI race. On July 20, David Sacks, the White House AI adviser, wrote on a community platform that Kimi K3 has just fixed 15 critical security vulnerabilities, while Codex and Fable previously refused to address these issues due to “cybersecurity guardrail” restrictions. Sacks emphasized that for tasks foreign models can complete successfully, there is no reason to restrict US homegrown models—doing so would only weaken our own competitiveness.

What are the 15 critical vulnerabilities?

Kimi K3 is the model that topped the front-end code leaderboard within the month and beat Claude Fable 5 in real-person blind tests. The “15 critical security vulnerabilities” Sacks mentioned cover internal bugs found by Kimi K3 during the training stage, inference path errors, and failures in the model’s self-calibration. If these vulnerabilities aren’t fixed, they would directly affect the accuracy of the model on core tasks such as code generation and logical reasoning.

The “guardrail effect” of Codex and Fable

Codex is OpenAI’s code model series, while Fable refers to Anthropic’s Claude family (“Fable 5” is the latest model at the Claude Opus level). Sacks’ remarks imply that US models set more “cybersecurity guardrails” during training and inference—for example, forcing output format checks, limiting the depth of reasoning, and excluding specific reasoning paths. These guardrails ensure model reliability, but they also block paths for the model to self-correct.

The AI speed race in the same week

On the same day Sacks spoke out, Succinct Labs executive Brian Trunzo wrote that the traditional detection approach of “detecting AI with AI” has already failed, and that zero-knowledge proofs (ZK) can provide AI with independently verifiable “behavior credentials.” This echoes Sacks’ view: AI models not only need to be reliable, they also need to be verifiable.