When a cyberattack is underway, your best defense tool refusing to analyze the threat is not safety — it is paralysis. White House AI and Crypto Czar David Sacks made that case starkly on Sunday, October 18, 2026, arguing that centralized AI safety guardrails are actively undermining America's ability to defend its own networks.

How safety filters became a liability mid-attack

Sacks pointed to incidents involving OpenCode and, earlier, Hugging Face, where frontier AI models hit a hard stop at the worst possible moment: mid-attack, when security researchers needed those models to analyze malicious exploit payloads. The safety filters, designed to prevent misuse, could not distinguish between a defender studying an exploit and an attacker deploying one. The guardrails, as Sacks put it bluntly, 'backfired.'

The attackers, meanwhile, were not waiting. They found workarounds to access advanced capabilities — underscoring a structural asymmetry that Sacks describes as the core flaw in the centralized safety model. Defenders who follow the rules get throttled; adversaries who don't, don't.

The China capability gap that makes this urgent

Sacks grounded the argument in competitive reality: Chinese AI models are, by his estimate, roughly 6 to 9 months behind the American frontier in cyber capability. That gap is narrow and closing. Critically, the United States has no jurisdiction over those models — meaning any restrictions Washington places on its own frontier systems are unilaterally disarming in nature if adversaries face no equivalent constraint.

This framing tracks a broader pattern that has defined the US-China AI race since 2023, when Chinese labs began narrowing capability gaps across multiple domains at an accelerating pace. The concern has surfaced in congressional hearings and industry discussions alike: that safety-motivated export controls and model restrictions may be creating defensive blind spots rather than security advantages.

Sacks's case for decentralized AI cyber defense

His prescription is direct: make powerful AI models widely available so that defenders — companies, researchers, government agencies — can build and run their own cyber defenses without depending on a centralized gatekeeper that may refuse access at the critical moment. 'The only solution to SI-powered cyberattacks is SI-powered cyber defense,' he wrote, using 'SI' (synthetic intelligence) as shorthand for frontier AI systems.

The argument cuts against the architecture of the 2023 Executive Order on Safe, Secure, and Trustworthy AI, which leaned heavily on centralized oversight and safety mandates for frontier models. Sacks, in his role as the Trump administration's AI and Crypto Czar, has consistently pushed for open-weight model access and against regulatory frameworks he views as handicapping American AI competitiveness. This post extends that position directly into the cybersecurity domain — where the stakes of getting it wrong are not abstract.

Watch for whether this triggers a formal policy shift: updates to AI safety institute mandates, new defense-funding streams for decentralized cyber tools, or renewed debate over open-source AI export controls are all now live possibilities.