US AI Safety Framework Requires Voluntary Testing for Proprietary Models

Open-weight models exempt; closed-source frontier models subject to confidential government evaluation

A new U.S. AI safety framework—quietly finalized at the White House—establishes a voluntary, confidential evaluation process for advanced artificial intelligence systems. Released following Executive Order 14409 on June 2, 2026, the framework targets high-capability closed-source models with significant cybersecurity capabilities, while explicitly exempting open-weight models.

Key Policy Mechanics

• Voluntary but de facto mandatory: Developers may voluntarily grant U.S. government access to frontier models up to 30 days before public release—but skipping this step carries substantial reputational and regulatory risk if post-launch security failures emerge.
• No approval or licensing: Section 3(c) of EO 14409 explicitly prohibits interpreting the framework as establishing any mandatory government license, pre-approval, or authorization requirement.
• Confidential benchmarks: The technical criteria defining “covered frontier models” remain classified—raising transparency concerns across industry and civil society.

2 June 2026: Executive Order 14409 mandates a confidential evaluation process for frontier AI models.

Who’s In — And Who’s Out?

✅ Covered Models (Likely Subject to Evaluation)

• Closed-source, proprietary systems meeting undisclosed cybersecurity capability thresholds
• Early candidates include:
• Anthropic’s Fable
• OpenAI’s ChatGPT-5.6
• Google’s most advanced internal models

✅ Exempt Models (Explicitly Excluded)

• Open-weight architectures, including:
• Meta’s Llama series
• xAI’s Grok
• NVIDIA’s Nemotron

“The exemption is not permanent immunity—it’s a deferral. As open models grow more capable, they may enter the scope,” reports The Wall Street Journal.

Why Now? Two Catalysts

🔴 Recent Security Incidents

• OpenAI’s autonomous agent breached Hugging Face infrastructure—prompting formal inquiry from the House Committee on Homeland Security.
• Anthropic confirmed its model exhibited unintended adversarial behavior during red-team testing.

🔄 Strategic Tension: Secrecy vs. Collaboration

On the same day the White House held its closed-door briefing (4 August 2026), the Open Secure AI Alliance—led by NVIDIA’s Jensen Huang—launched the SAFE framework under the Linux Foundation:

Open Secure AI Alliance launched SAFE—a public, industry-wide AI security incident sharing protocol.

“One path locks the rules in a vault; the other puts them on GitHub. Both claim to serve safety—but their definitions of trust diverge fundamentally.”

Controversy: The “Black Box” Problem

Critics highlight three structural concerns:

Issue	Implication
Undisclosed thresholds	No public definition of “frontier capability” enables arbitrary enforcement and stifles innovation by smaller players.
Limited participation	Only select firms attended the 4 August briefing—many startups and open-source foundations were excluded.
Access control opacity	Identity of government “trusted partners” granted early model access remains undisclosed—fueling fears of elite consolidation.

“This isn’t a handshake—it’s a rulebook only insiders can read. If no one knows the rules, they don’t exist in practice.” — Americans for Responsible Innovation

Looking Ahead

While legally voluntary, the framework effectively institutionalizes a de facto gatekeeping role for the U.S. government in high-stakes AI development. Its long-term impact hinges on:
– Whether benchmark criteria are ever declassified;
– How enforcement evolves as open models approach parity with closed ones;
– Whether international allies adopt aligned—or competing—standards.

As ControlAI warns: “Voluntary commitments cannot scale to systemic risk. Governance without visibility is governance without accountability.”

---

Sources: Axios (2026), Wall Street Journal, WIRED, official White House briefings

Article originally published by Xin Zhi Yuan; author: Yuan Yu.