The Big Picture

The same administration that spent a year dismantling AI oversight is now quietly drafting rules — but the framework being built hands the Pentagon and intelligence agencies first access to the most powerful models on earth, not the public. Meanwhile, the digital infrastructure billions of people trust to keep them safe was compromised through a screensaver, a ransomware gang had a 36-day head start inside hospital firewalls, and the browser tool journalists and dissidents rely on to stay alive was silently broken. Power is consolidating. Accountability infrastructure is cracking. The people who pay the price are never the ones in the briefing room.

Today's Stories

The White House Wants to "Vet" AI — But It's the Pentagon That Gets First Look

The Trump administration is reportedly drafting an executive order requiring government oversight of frontier AI models before public release. Sounds like accountability. Read the fine print. According to The New York Times, senior officials briefed Anthropic, Google, and OpenAI executives last week. Bloomberg's follow-on reporting identifies the candidate agencies as the NSA, the White House Office of the National Cyber Director, and the director of national intelligence. This is not a consumer safety framework. It is a national security capture play — giving military and intelligence agencies first access to evaluate models for "security vulnerabilities and military applications" before anyone else sees them. Anthropic's unreleased Mythos model, reportedly capable of autonomous network intrusion, is cited as a catalyst. The administration that called regulation a threat to innovation didn't change its mind. It found a way to route the most powerful AI tools through the security state first. Open-source models — once distributed, impossible to recall — face a separate existential compliance question this framework cannot answer. Watch who actually benefits from "safety."

Ransomware Had a 36-Day Head Start on Hospital Firewalls. Patients Paid.

The Interlock ransomware group was inside Cisco's Secure Firewall Management Center for 36 days before Cisco even published a patch, exploiting CVE-2026-20131 — a critical flaw that granted unauthenticated root access — since January 26, according to Amazon's threat intelligence team. Cisco didn't disclose until March 4. Confirmed victims include DaVita, a U.S. dialysis provider; the Kettering Health hospital network; Texas Tech University; and the city of Saint Paul, Minnesota. Interlock paired encryption with data theft and explicit GDPR threats to maximize pressure on institutions that cannot afford downtime — institutions that serve working-class patients who have no alternative. Organizations that applied Cisco's patch the day it dropped may still be compromised from January. The lesson for regulators and the public: critical healthcare infrastructure runs on commercial software from vendors who disclosed a critical vulnerability only after attackers had already been inside for over a month. That is not a bug. That is a business model.

The Internet's Trust Layer Was Broken Through a Screensaver

DigiCert — one of the largest Certificate Authorities on earth, whose digital signatures tell your computer what software is safe to install — confirmed it was breached through a malicious screensaver file sent via a Salesforce customer support chat. The attacker tried five times; four were blocked. The fifth succeeded because one analyst's security agent was misconfigured. DigiCert didn't discover a second compromised machine until ten days later. During that window, the attacker extracted initialization codes for approved Extended Validation code-signing certificates and used them to sign Zhong Stealer malware, a credential and cryptocurrency stealer linked to a Chinese e-crime group. DigiCert revoked 60 certificates — 27 tied to the attacker — but only acknowledged the breach publicly after abused certificates were spotted in the wild. Then a faulty Microsoft Defender update began removing legitimate DigiCert root certificates from Windows systems entirely. The infrastructure the entire internet uses to establish trust is held together by a single misconfigured endpoint and a screensaver. That should terrify everyone.

Tor's "New Identity" Button Wasn't Working. Journalists and Dissidents Didn't Know.

Firefox's private browsing mode was leaking a stable identifier — derived from the order of IndexedDB database entries — that persisted across private windows as long as the browser process ran. In Tor Browser, it survived the "New Identity" feature: the reset button that journalists, dissidents, and activists rely on as a clean break between sessions. Any website could observe the ordering, derive the fingerprint, and link sessions that were supposed to be anonymous. Mozilla patched it as CVE-2026-6770 in Firefox 150 and ESR 140.10 on April 21. The Tor Project released Tor Browser 15.0.10. But the people who needed this guarantee most — the ones whose lives can depend on it — were using a broken tool without knowing it. If you haven't updated, every Tor session opened in the same browser process can be linked together by any site you visit. Update now.

AI Labs Are Becoming Consulting Firms — and Wall Street Is Buying In

OpenAI and Anthropic are no longer content selling API access. Both have launched consulting joint ventures to capture enterprise implementation revenue directly. OpenAI's "Deployment Company" — backed by TPG, Brookfield, Advent, and Bain Capital — has raised roughly $4 billion at a $10 billion pre-money valuation. Anthropic's parallel JV with Blackstone, Hellman & Friedman, and Goldman Sachs is funded with $1.5 billion. Anthropic simultaneously disclosed that finance is its second-highest revenue segment, and shipped ten ready-made agent templates for financial services — pitchbook generation, KYC screening, month-end close — integrated with FactSet, S&P Global, and Morningstar. Anthropic CEO Dario Amodei is reportedly in conversations with JPMorgan's Jamie Dimon. The labs built the technology. Now they want the consulting margins too. The workers whose jobs these agents are designed to automate are not in those boardrooms. The private equity firms funding the rollout are.

What to Watch

• [CONFIRMED] If the White House executive order on AI model vetting is signed, then the NSA and intelligence community gain structural first-access to frontier AI systems — with no mandated public transparency or consumer safety review built into the reported framework. (Confirmed: based on NYT and Bloomberg reporting on briefings already conducted)
• [ASSESSED] If Interlock's 36-day pre-patch exploitation of Cisco firewall infrastructure follows the group's established pattern of targeting healthcare and public institutions, then additional hospital networks and municipal governments that applied the March 4 patch without conducting a retrospective compromise assessment remain actively breached. (Assessed: based on confirmed victim list and exploitation timeline)
• [ASSESSED] If OpenAI and Anthropic's consulting joint ventures successfully convert enterprise deployments into recurring services revenue, then traditional labor in financial services, legal, and IT modernization faces displacement pressure from firms that also own the underlying models — with no worker representation in the deal structures reported. (Assessed: based on disclosed JV structure and Anthropic's finance agent product launch)
• [SPECULATIVE] If the White House AI vetting framework effectively excludes open-weight models — which cannot be recalled once distributed — then the practical result is a regulatory structure that entrenches closed, corporate-controlled AI while foreclosing community and public-interest alternatives. (Speculative: structural implication of reported framework, mechanism not yet confirmed)

The Closer

The screensaver that cracked DigiCert, the 36-day ransomware window inside hospital firewalls, the broken Tor reset button — none of these failures happened to the executives who will testify about them. They happened to patients, to activists, to the people who had no choice but to trust infrastructure they had no power to inspect. That is the through-line this week: the systems that govern our digital lives are built for the convenience of the powerful and secured, when they are secured at all, as an afterthought. Accountability journalism doesn't just document the breach. It asks who decided the screensaver filter wasn't worth the budget.