What the Pope Actually Said About AI

May 26, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 26, 2026

HEADLINES

Project Glasswing Update

At the end of last week, Anthropic published an update on Project Glasswing — the initiative through which Mythos was rolled out in limited form to around fifty partners. The headline number: more than ten thousand software vulnerabilities of high or critical severity identified so far, with partners averaging hundreds of severe finds in just over a month and a false positive rate under 10%. Mozilla alone found and fixed 271 vulnerabilities — more than ten times what they found using Opus 4.6. One firm in finance used Mythos to catch and block a $1.5M wire transfer to a threat actor in real time.

But the bigger story is where the bottleneck now sits. Mythos has fundamentally changed how fast vulnerabilities are found — the problem is that patching hasn't kept pace. Anthropic reports that some open-source maintainers have actually asked them to slow down disclosures because they can't design patches fast enough. Box CEO Aaron Levie called it a microcosm of AI deployment broadly: engineers don't go away, the constraint just moves. "Far from AI magically solving all of this, there is still major triage work and human judgment required. As a result, we're about to enter a security engineer boom. Jevons paradox all over again." Meanwhile, Anthropic confirmed they're working to expand Glasswing to allied governments, though a public release of Mythos-class models remains on an unclear timeline.

White House Approves $9B for Spy Agency AI

The White House has quietly approved a secret $9 billion budget request from the CIA and NSA to build their own inference cluster — because they currently can't run the latest AI models on classified systems. The funding is earmarked for Nvidia Blackwell chips and supporting infrastructure. Separately, sources told The Information that White House Chief of Staff Susie Wiles has approved Anthropic model access at the NSA despite the Pentagon having designated Anthropic a supply chain risk — and that a formal contract is being finalized. The distinction that makes this work: unlike the Pentagon, the NSA and CIA are legally prohibited from domestic surveillance, making it easier for them to accept Anthropic's guardrails on that front.

DeepSeek Makes 75% Discount Permanent, Eyes $10B Funding Round

DeepSeek has made their token discount permanent — meaning a model roughly equivalent to Opus 4.5 now costs 44 cents per million input tokens and 87 cents per million output tokens, while Anthropic and OpenAI have actually raised prices on their newest models. That's one-seventh the cost of Opus 4.6 at launch. Simultaneously, DeepSeek is closing in on a $10B funding round — its first outside capital — at a reported $45B valuation, with participation expected from the Chinese Government's AI Industry Investment Fund, Tencent, and JD.com. Founder Liang Wenfeng has told investors the focus remains on open-source research and AGI, not monetization. Bloomberg analysts see this as part of a broader decoupling: Asia treating AI tokens as tradable industrial commodities rather than premium software.

Grok V9-Medium Finishes Training; Grok Build Expands to Pro Users

Despite xAI's pivot toward becoming an AI cloud company, it isn't done training its own models. Elon Musk announced that Grok V9-Medium — a 1.5 trillion parameter foundation model, likely in the same class as Mythos or GPT-5.5 — has completed training, with fine-tuning underway and a public release expected in two to three weeks. Separately, Grok Build, xAI's Claude Code competitor, has rolled out to SuperGrok and X Premium subscribers, previously restricted to the $300/month Grok Heavy tier. Early impressions have been positive.

MAIN STORY

What the Pope Actually Said About AI

Pope Leo XIV's first encyclical, Magnifica Humanitas — subtitled "On Safeguarding the Human Person in the Time of Artificial Intelligence" — generated more serious AI discourse in twenty-four hours than most news cycles manage in a month. The core recommendation before diving in: try to actually read it. Most of the takes, from all directions, were bad. The document is not the AI doomer manifesto some wanted, nor the techno-optimist endorsement others wished for. It's a flag-planting document that lays out the foundational principles Leo believes must guide how we navigate this moment — and the central argument, more than anything else, is that human value cannot be reduced to an intelligence benchmark.

Background: Who Is Leo, and What Is an Encyclical?
The first American pope, elected to address the AI age.
An encyclical is a major teaching letter from the Pope — a tradition going back to the early church — that interprets existing Catholic teaching in light of new circumstances. Leo XIV chose his name as a direct tribute to Leo XIII, who authored Rerum Novarum in 1891 to address labor, wages, and industrial society. Leo XIV made AI the central social question of his papacy from his very first address. He's also notably of the modern world in a way his predecessors weren't — he has a cell phone, computer, Apple Watch, and watches cable TV. His read on AI throughout has been less about designating it good or bad, and more about acknowledging it as a pivot point in both history and faith.

The Encyclical Itself
42,300 words; more exploration than condemnation.
The document positions us at "the beginning of the beginning." Its warnings are prospective — about what could happen — rather than indictments of what is happening now. Christopher Hale put it well: "Anyone who thinks Pope Leo XIV is an AI doomer either hasn't read the encyclical in full or doesn't understand Catholic theology. The Pope is an AI realist. He knows its growth is inevitable. He just wants to ensure it's always in service of the human person." The central plea, as much as one can reduce 42,000 words, is that humans — not markets, not AI systems — remain the barometer of success. The much-quoted line about profits not justifying sacrificing jobs is not an argument against technological change; it's a restatement of a tension that has existed since the Industrial Revolution.

The Bad Takes
A Rorschach test, in both directions.
The discourse split cleanly into two failure modes. On one side: people who ignored the text entirely to litigate the optics of Anthropic being invited to speak at the ceremony. On the other: people who cherry-picked a single quote to validate whatever they already believed. The jobs quote — "The pursuit of greater profits cannot justify choices that systematically sacrifice jobs" — drew responses like Blake Scholl's: "Bad take from the Pope. Tech revolutions tend to eliminate some jobs while creating others." Senator Chris Murphy and author Karen Hao went the opposite direction, reading into the encyclical a full indictment of AI's current supply chain and extractive practices that the text doesn't actually make.

Dean Ball and the AI Cognition Debate
The most revealing frustration, and the most revealing response to it.
Dean Ball's critique was the most intellectually interesting of the bad takes: that the encyclical punts on the central challenge by axiomatically denying AI cognition rather than grappling with what humanity should do as we are "eclipsed as the smartest entities on the planet." Jonathan Liedl's response captured the counterpoint well: "Guy is extremely mad that the Pope of the Catholic Church believes in the soul, apparently." Ball clarified he wasn't asking the Pope to ensoul AI — he wanted the Church to help us understand what humans should do in the face of that eclipse. That question, more than any other, actually points toward what the encyclical is doing.

Catholics Have the Clearest Read
"The Pope is an AI realist. He knows its growth is inevitable."
Fascinatingly, Catholics on the whole had the sharpest read of the document's actual tone. Christopher Hale: "Anyone who thinks Pope Leo XIV is an AI doomer either hasn't read the encyclical in full or doesn't understand Catholic theology. The Pope is an AI realist. He knows its growth is inevitable. He just wants to ensure it's always in service of the human person." The Catholic Sat account put the TLDR well: "AI is not inherently evil, but is never neutral and carries risk of power concentration, inequality, and loss of human dignity. The church offers principles for discernment rather than blanket rejection." That's the right read. It's a document about the beginning of the beginning — warnings for what could be, not an indictment of what is.

The Core Plea: Humans as the Barometer
Markets are supposed to serve people — AI is reigniting a very old tension.
The encyclical's subtitle is "Safeguarding the Human Person in the Time of Artificial Intelligence," and that's the core plea: keep real living humans — not AI systems, not anonymous market forces — as the measure of success. The jobs quote that got so much attention isn't an argument against technological change. It's a restatement of the fundamental tension between markets as efficient allocators of scarce resources and markets as ultimately servants of human wants and needs. AI is reigniting that tension at massive scale, but the tension itself is nothing new. Similarly, the colonialism argument isn't a declaration that AI is currently exploitative — it's a warning about what happens when a new resource (data, specifically health data) becomes concentrated power. Paragraph 178: those who control the health data of entire peoples "possess a structural leverage over the future, for they can shape needs and markets."

The Central Argument: Human Value Cannot Be Benchmarked
A flag planted before the real debate begins.
The deepest argument in the document — and the one that will age most interestingly — is that human value cannot be reduced to an intelligence benchmark. Even if and when AI surpasses human cognitive performance, it remains categorically different. Paragraph 99 draws the line: these systems process data; they do not undergo experiences, feel joy or pain, mature through relationships, or bear moral responsibility. They can simulate empathy and imitate language, but "they do not understand what they produce." Closely related is paragraph 118, which reframes human limitation itself — illness, vulnerability, suffering — not as defects to be optimized away, but as conditions through which humanity matures. This may not be a controversial take for most people who use AI every day — but as AI gets more advanced, more people will have more questions about it. In that sense, Magnifica Humanitas is a flag-planting: preparation for the period when that conversation accelerates.