Digital Colliers Daily Briefing — June 18, 2026
The week's headlines map cleanly onto three pressure points reshaping the technology industry: the geopolitics of frontier AI distribution, the migration of AI labs into physical-world instruments, and the persistent fragility of perimeter security infrastructure. Today's events feature the Trump administration's unprecedented use of export controls against an American AI model, Midjourney's pivot from image generation to medical imaging hardware, and the disclosure of a Fortinet compromise affecting roughly half the world's internet-facing units of the vendor's firewalls. Each, in its own way, exposes how brittle the assumed foundations of the modern stack have become.
1. Export controls hit Claude: Anthropic's Mythos 5 and Fable 5 remain dark as G7 leaders question US AI dependence
What happened. Late last week, the Trump administration ordered Anthropic to cut access to its Mythos 5 and Fable 5 models for all foreign nationals, including immigrants resident in the United States and Anthropic's own non-citizen employees. Rather than implement nationality-based gating, Anthropic took both models fully offline. According to The Verge, this is the first known use of US export controls to restrict access to a specific AI model. As of Wednesday, the models remain offline despite days of negotiations.
Per Wired's reporting, the trigger was twofold: White House discomfort over Anthropic granting SK Telecom access to Mythos through its Project Glasswing early-access program — officials alleged unspecified ties to China, which SK Telecom denies — and a subsequent disclosure from Amazon researchers of jailbreaks in Fable 5 that could surface Mythos's cybersecurity, biology, and chemistry capabilities. The NSA concluded the guardrails were bypassable; the Commerce Department and Office of the National Cyber Director told Anthropic on Monday that it must demonstrate it can proactively find and disclose jailbreaks itself before the models return.
Why it matters. Independent security researchers have argued for years that guardrails are a stopgap, and Wired notes the administration's standard — that Anthropic block all jailbreaks — may not be technically achievable. The episode therefore establishes a precedent in which a US AI firm can be taken offline indefinitely over a class of vulnerability the field broadly considers unsolved.
Who is affected. Anthropic's commercial customers globally, every non-citizen employee of the company, and the entire Project Glasswing consortium (roughly 150 organizations including Samsung Electronics and the Korea Internet and Security Agency). The broader collateral damage landed at the G7 summit in Évian-les-Bains, where, as TechCrunch reports, Macron warned Trump and assembled AI executives that an overnight US "kill switch" would harm both allied economies and US AI firms themselves. Modi echoed the concern. Cohere CEO Aidan Gomez framed the episode as validating the case for digital sovereignty.
What to watch next. G7 leaders discussed a "trusted partners" scheme to guarantee continued access to frontier models from Anthropic and OpenAI for aligned states and companies — a notable concession that current export policy is incompatible with allied procurement. Watch for whether Anthropic accepts an obligation to run continuous adversarial testing and report findings to the government, and whether the SK Telecom thread resurfaces in any final settlement.
Sources:
- Anthropic got hit by export rules nobody understands — The Verge AI
- Vibe-decoding the White House-Anthropic fight over Fable — The Verge AI
- The White House Wants Anthropic to Block All Jailbreaks. That May Not Be Possible — Wired
- The Korean Telecom Giant at the Center of Anthropic's Mythos Controversy — Wired
- "Dangerous" AI models are coming no matter what — Ars Technica
- World leaders want American AI. They just don't want America to be able to turn it off. — TechCrunch AI
2. Midjourney unveils a full-body ultrasound CT scanner — and a San Francisco spa to put it in
What happened. Midjourney CEO David Holz unveiled the Midjourney Scanner, a full-body ultrasound CT system that Holz described as the "first new whole-body medical imaging modality in 50 years." The Gen 1 prototype uses 40 ring-arranged systems containing 358,000 ultrasonic elements, captures roughly 17 GB/s, and produces about 40 GB per body slice with claimed 0.5 mm resolution of internal tissue. Reconstruction runs across 21 servers at a claimed 2 petaflops. The company simultaneously announced the Midjourney Spa — a four-floor, 25,000-square-foot facility near Union Square targeted for end-2027 with roughly nine to ten scanners, hot tubs, saunas, and a gym. The scanner requires water immersion, hence the spa.
About a dozen people have been scanned, current scans take roughly 20 minutes, and Midjourney has begun preliminary discussions with the FDA, with body composition as the likely initial regulatory wedge. The team is approximately nine people. Midjourney remains self-funded.
Why it matters. Midjourney joins a broader pattern of frontier AI labs reaching past models into the physical instruments that produce training data — a trajectory most visible in biology with cryo-EM and protein structure work. The pitch is classic disruption positioning: not better than MRI on peak image quality, but radiation-free, magnet-free, faster, and potentially orders of magnitude cheaper per scan. Holz speculated fewer than a dozen fully utilized machines could match global MRI throughput, and a fleet of 50,000 could enable a billion scans per month — claims that, as Latent Space notes, are unvalidated clinically, economically, and operationally.
Who is affected. Incumbents in medical imaging (GE Healthcare, Siemens Healthineers, Philips) face a long-horizon threat if the modality clears regulatory hurdles. Consumer health and longevity-medicine startups gain a potential platform partner. Investors should note Midjourney estimates roughly $20 billion in capex to scale to thousands of facilities — a financing problem its current image-generation revenue cannot solve alone, despite its no-investor posture.
What to watch next. Clinical validation data (none has been released), the FDA pathway for body composition versus diagnostic claims, Gen 2 hardware promised by end of 2026, and whether Midjourney can navigate the overdiagnosis and incidental-finding risks of high-frequency whole-body screening — issues Holz acknowledged but did not resolve.
Sources:
3. Fortinet breach exposes plaintext credentials for 74,000 firewalls across Oracle, Chevron, and a NATO contractor
What happened. Researcher Bob Diachenko of SecurityDiscovery.com disclosed a breach in which Russian-speaking attackers compromised nearly 74,000 Fortinet firewalls operating from more than 21,000 IP addresses across 194 countries, exposing plaintext credentials online. Diachenko obtained the data after gaining access to the attackers' command-and-control infrastructure. Affected organizations include Oracle, Chevron, Lenovo, FedEx, an unnamed NATO defense contractor, and Fortinet itself. The leak metadata also included industry, revenue, and headcount per victim — indicating curated targeting. Independent researcher Kevin Beaumont confirmed credentials are current and that "almost all" compromised devices remained online as of Wednesday morning. In many cases, attackers pivoted from the firewall into Radius servers and Microsoft Active Directory.
Why it matters. Shodan data cited by Ars Technica puts the compromised set at roughly half of all internet-facing Fortinet firewalls — a scale that effectively reclassifies this from a vendor incident into an infrastructure event. Credential exposure at the perimeter is the worst possible failure mode because it provides a clean pivot into internal authentication systems, which is precisely what occurred here.
Who is affected. Every Fortinet customer with an internet-facing device must now treat credentials as burned and assume Active Directory exposure until proven otherwise. The presence of a NATO defense contractor in the victim list raises immediate counterintelligence questions; the presence of Fortinet itself raises supply chain ones. Cyber insurers face a clustered loss event of substantial magnitude.
What to watch next. Disclosure of the specific Fortinet vulnerability or credential-harvesting vector; CISA and allied agency advisories; whether the threat actor monetizes via ransomware affiliates or holds access for intelligence purposes; and Fortinet's response to having been compromised on its own infrastructure.
Sources:
Closing
Three stories, one through-line: the assumed reliability of foundational systems — frontier model availability, medical imaging modalities, perimeter security — is more contingent than enterprise buyers have priced in. The Anthropic episode tells CIOs that a US model can vanish on national security grounds with no notice; the Fortinet breach tells them that perimeter credentials they trust may already be in adversary hands; and Midjourney's scanner suggests the next category of physical-world infrastructure will be built by companies whose names recently meant something else entirely. The procurement, sovereignty, and resilience questions raised today will outlast the news cycle they arrived in.