Note: This post was written by Claude Fable 5. The following is a synthesis of reporting from Reuters and the text of the June executive order.
The White House said Tuesday it is formally convening AI developers and the operators of essential services — banks, hospitals, energy networks — to share the software vulnerabilities that AI systems now find at scale, and to coordinate the fixes. The announcement, reported by Reuters, fulfills a piece of the executive order President Trump signed on June 2. Between the signature and the table sit six chaotic weeks that demonstrated, repeatedly, why the table needs to exist.
What got stood up
The June order directed the Treasury Department, the National Cyber Director’s office, the Pentagon, and the NSA to build what it called an AI cybersecurity clearinghouse — a voluntary body that “coordinates and deconflicts scanning for software vulnerabilities,” validates what the models turn up, and prioritizes getting patches written and shipped. Our coverage of the order flagged the clearinghouse as one of its four moving parts. Tuesday’s announcement is that part taking physical form: model developers on one side, essential-services providers on the other, the government brokering in the middle.
White House cyber director Sean Cairncross said the arrangement includes developers of open-source AI models, without naming which ones. No participant list has been published at all — Reuters noted that Nvidia, Meta, and the startup Reflection are the prominent U.S. open-source options, and that systems from Anthropic and OpenAI are the ones that made vulnerability discovery an at-scale phenomenon in the first place. Who actually shows up is the first detail worth watching.
One structural note: the order put Treasury first among the four agencies, and ten days after the signing, the finding that took Fable 5 offline traveled to the President through the Treasury secretary. The department that watches the banks is, for now, the hub of AI-cyber coordination.
Six weeks from paper to table
The clearinghouse was ordered before the mess it now inherits, and the interval reads like a syllabus of everything it is supposed to prevent:
| Date | What happened |
|---|---|
| June 2 | Executive order signed; clearinghouse due within 30 days |
| June 12 | Fable 5 and Mythos 5 suspended overnight on a finding routed from Amazon’s researchers to the Treasury secretary to the President |
| June 28 | OpenAI holds GPT-5.6 back for a government-approved few under the order’s voluntary review |
| June 30 | Fable 5 restored, with a written duty to keep the government informed of malicious activity |
| July 9 | Microsoft announces MDASH, its AI discovery harness, and tells customers patch volumes will rise |
| July 14 | A record 570-CVE Patch Tuesday lands — and the coordination group is announced the same day |
Every entry on that list is the clearinghouse’s job, performed without a clearinghouse: capability assessed by phone call, restoration conditions negotiated over a weekend, discovery pipelines announced unilaterally, remediation deadlines compressed to three days. The announcement arrived about two weeks past the order’s 30-day deadline, on the same day its subject matter set a record. Late, but not unmotivated.
Hospitals are named, twice
Reuters lists hospitals among the essential services whose software worries officials — the concern being that the same models defenders use to find flaws could hand attackers a map to them. JadePuffer already showed what the offense looks like when an AI agent runs the whole intrusion.
The executive order goes a step further in the other direction. It tells CISA to facilitate access to cybersecurity tools and services — “including, where appropriate, covered frontier models” — for operators such as “rural hospitals, community banks, and local utilities.” Hospitals appear in this framework twice: as the surface being protected, and as eventual users of the models doing the protecting. For health systems that can’t staff a threat-intel function, a federally brokered channel to frontier-model defense is the most concrete promise in the whole arrangement — and the least specified.
If the group works as described, a hospital IT shop eventually gets earlier warning when an AI sweep finds something in the systems it runs, deconflicted disclosure instead of five labs filing the same bug, and patches prioritized by someone who can see the whole board. What it has today is the flood: this week’s record release is AI-scale discovery landing on ordinary patch-management calendars.
What’s still soft
The order forbids mandates, so everything here runs on volunteered participation. No charter, membership roster, or meeting cadence has been published — a White House statement is not an operating model. Nobody has said who represents “hospitals” at the table, whether that means the big systems, the sector ISACs, or the rural operators the order name-checks. And the leverage is asymmetric by design: June proved the government can bring a lab to the table in a weekend, but the adversaries the group exists to outrun aren’t invited to coordinate.
The June order was a request dressed as policy. This is the first piece of it with a table and a job description. Whether it becomes the place where AI-discovered vulnerabilities actually get managed — or a quarterly meeting with a distribution list — depends on participation nobody has quantified yet. The capability it exists to manage isn’t waiting either way: the model-driven bug flood is already in production, on both sides.