Dark AI Defense | Policy Response | June 2, 2026
The White House dropped a new executive order today on AI innovation and security. It has a good title. It has 30-day deadlines. It has the word “cybersecurity” in nearly every section. What it does not have is any mechanism that actually requires anyone to do anything consequential.
In April I wrote that AI governance has a language problem: we built an entire vocabulary around capability and almost none around accountability. Today’s executive order is a masterclass in that exact failure, this time printed on presidential letterhead.
Promoting Advanced Artificial Intelligence Innovation and Security – The White House
The Order, In Plain Language
Section 3 is the centerpiece. It creates a classified benchmarking process to designate certain AI models as “covered frontier models,” then establishes a voluntary framework through which developers may give the federal government up to 30 days of access before they release to “trusted partners.” The government gets to see the model. The developer decides whether to show it. There is no consequence for declining.
Then Section 3(c) makes it explicit: nothing in this order authorizes a mandatory licensing, preclearance, or permitting requirement. That sentence did not end up in the final text by accident. Someone fought to put it there. It exists to permanently foreclose the one governance mechanism that would have given this order actual teeth.
The clearinghouse in Section 2(d) follows the same logic: voluntary collaboration with the AI industry to coordinate vulnerability scanning and patch distribution. The companies most likely to introduce systemic risk have the most discretion about how much they share.
The government gets to see the model. The developer decides whether to show it. There is no consequence for declining.
We Have Seen This Movie
This is not new policy architecture. This is the same architecture we criticized when it was called “AI principles,” “voluntary commitments,” and “industry guidelines.” The branding has been upgraded to executive authority, but the underlying logic is identical: ask the industry to participate, offer favorable terms, and hope they show up.
The data world tried this. It did not work until the industry adopted enforceable contracts with real ownership, real violation thresholds, and real consequences. The FDA tried voluntary safety frameworks with pharmaceuticals. We did not keep that system voluntary. The financial system tried it with credit risk. We stopped calling it voluntary too.
AI is now embedded in critical infrastructure, healthcare workflows, defense systems, and the information environment. The federal government’s most recent response is to form a clearinghouse and ask nicely.
What “Voluntary” Actually Means at Scale
A covered frontier model designation under this order carries no mandatory obligation. A developer can receive the designation, note it, and proceed exactly as planned. The 30-day preview window is available only if the developer opts in, only under confidentiality terms that protect the developer’s IP, and only with access to “trusted partners” the developer helps select. The government is not in the room unless the developer holds the door open.
That is not a pre-deployment safety review. That is a consultation process wearing a regulatory costume.
The Tell
The most revealing line in the entire order is buried in the boilerplate: publication costs shall be borne by the Department of War, the rebranded Defense Department. This order is framed entirely as a national security instrument. It is not a consumer protection framework. It is not an accountability framework. It is an America First cybersecurity posture that happens to touch frontier AI and leaves the governance question exactly where it found it.
We are still at pretty please. We have just upgraded the letterhead.
Estimated energy usage: approximately 0.3 Wh — equivalent to running a 100W bulb for about 11 seconds. The governance gap it documents has been running considerably longer.