The Order That Wasn’t: Trump Pulls His Own AI Security Rule Hours Before Signing
President Donald Trump was scheduled to sign an executive order on AI security on Wednesday. He didn’t.
Hours before a planned signing ceremony, Trump postponed the order indefinitely, telling reporters: “I don’t want to get in the way of that leading.” The order would have created a voluntary 90-day testing and vetting regime for frontier AI models.
Yes, the same Donald Trump who tore up Biden’s AI safety executive order on day one of his second term, calling it “government overreach,” then wrote his own version, then shelved it too.
In five months, the administration has gone from “anything goes” to “please show us your models” to “actually, never mind.” The whiplash is remarkable even by Trump standards.
What the Order Actually Does
According to Reuters, which obtained the draft, the order establishes a voluntary framework where AI developers — OpenAI, Anthropic, Google DeepMind, Meta, and Microsoft — would:
- Provide pre-release access to powerful models to federal cybersecurity agencies (CISA, NSA’s Cybersecurity Collaboration Center, the FBI) 90 days before public launch
- Grant critical-infrastructure operators — banks, energy companies, telecoms, healthcare networks — early access to stress-test their defenses against new frontier models
- Participate without civil or criminal penalties for declining, though the government reserves the right to publicly name non-participants
The structure is calibrated against the specific cyber-threat surface that Anthropic’s Mythos exposed. Anthropic had already committed to sharing Mythos findings with partner governments; this order codifies that pattern as domestic policy.
The Bannon Factor: Why It Was Written — And Why It Was Pulled
Here’s where it gets interesting. The pressure driving this order doesn’t come from the usual AI safety lobby. It comes from Steve Bannon and right-wing organiser Amy Kremer, who have been lobbying the White House hard for mandatory security review of frontier models.
Bannon’s framing: the launches of Anthropic’s Mythos and OpenAI’s GPT-5.5-Cyber have shifted the cyber-threat surface in a direction the federal government cannot afford to ignore. This is national security, not tech regulation — or at least that’s how the MAGA wing is selling it.
Inside the West Wing, it’s a compromise. David Sacks (the crypto-optimist, pro-innovation AI czar) pushed back against mandatory disclosure on competitiveness grounds. The Bannon-Kremer national security camp pushed for hard requirements. The voluntary 90-day window was the middle ground.
But then Trump met with tech executives. According to TechCrunch and Bloomberg, the President was persuaded that the order’s language “could have been a blocker” for the AI industry. The pro-innovation wing won this round. The signing ceremony was scrubbed.
What’s Missing
A lot. The order:
- Doesn’t impose penalties for non-compliance — it’s voluntary
- Doesn’t define which specific models qualify as “frontier”
- Doesn’t name the federal-agency coordinating body that will run the framework
- Doesn’t specify which critical-infrastructure operators get early access
- Doesn’t set a timeline for the first model disclosed under the framework
The White House has signalled all five major labs will participate from launch, but “signalled” isn’t “mandated.” The first real test will be whether any model actually gets disclosed under this framework before Q3 2026 — which is the administration’s own suggested timeline.
The Bigger Picture: US vs. China vs. Itself
The order was supposed to arrive in a politically charged window. The recent Trump-Xi Beijing summit on AI guardrails and Nvidia H200 export licensing established a bilateral track on AI governance. This executive order was the domestic complement — showing the US can police its own frontier labs while demanding China do the same.
Instead, the US showed it can’t even police itself. The administration that gutted Biden’s order for being “anti-innovation” wrote its own softer version, then couldn’t even sign that.
But the order isn’t dead. The White House says it will be “refined and resubmitted.” The Bannon wing hasn’t gone away. The cyber-threats from Mythos-class models haven’t diminished. This is a delay, not a burial — and the question is what language gets changed to satisfy both the pro-innovation camp and the national security hawks.
What This Means for NZ
New Zealand has no equivalent framework. If the US establishes a pre-release disclosure norm for frontier models, it becomes a de facto global standard — one that NZ’s AI policy will eventually need to match or explain why it hasn’t. The EU’s AI Act already requires similar disclosures for high-risk systems. The US, EU, and China are all now building oversight regimes, and NZ is nowhere in that conversation.
❓ Frequently Asked Questions
Q: Is this the same as Biden’s AI executive order? Not quite. Biden’s order required mandatory safety testing and disclosure. Trump’s is voluntary, with no penalties for non-compliance. But the substance — pre-release government access to frontier models — is remarkably similar.
Q: Which AI companies are affected? The five expected to participate: OpenAI, Anthropic, Google DeepMind, Meta, and Microsoft. The order targets “frontier models” but doesn’t define the threshold yet.
Q: What happens if a company refuses to participate? Nothing legally — the order has no enforcement mechanism. But the government can publicly name non-participants, which creates reputational pressure.
Q: Does this affect AI models available in NZ? Indirectly yes. If US labs are required to share models with US agencies before release, it changes the timeline and transparency for global launches. NZ benefits from any safety review, even if we’re not at the table.
🔍 THE BOTTOM LINE
Trump killed Biden’s AI safety order, wrote his own softer version, then shelved that too — all in five months. The delay shows the pro-innovation wing won this round, but the national security pressure from Bannon hasn’t gone away. The order will return, probably with even fewer teeth. The US still has no binding AI oversight framework.
SOURCES
- The Next Web — Trump to sign AI oversight order
- Reuters — Draft executive order details
- U.S. News & World Report — MAGA-base pressure analysis
- The Register — Trump jumps from ‘anything goes’ to ‘strict regulation’
- Vorys — Legal analysis of White House AI governance plan
