The Day the Pope and an Atheist AI Founder Stood Together at the Vatican (While Anthropic Plans to Release “Too Dangerous” Mythos)

May 26, 2026 — Day 126 of our AI crisis timeline.

Yesterday at the Vatican, something historically unprecedented happened. Pope Leo XIV became the first pope ever to personally present an encyclical to the world — a role traditionally delegated to cardinals. The 83-page document, titled “Magnifica Humanitas” (Magnificent Humanity), focuses entirely on artificial intelligence.

But here’s the part that made global headlines: Seated among the cardinals and theologians to address a packed Vatican auditorium was Chris Olah — Anthropic co-founder, 33 years old, Canadian billionaire, and atheist. Together, the pope and the tech executive made what observers called “an unlikely duo” championing a partnership between the Catholic Church and Silicon Valley to develop AI safeguards.

Same day, buried in a paragraph near the end of an Anthropic blog post about “Project Glasswing” (the restricted-access program for their Mythos bug-finding AI), the company revealed they plan to “one day release models that match the performance of Mythos to the public, once it can make them safe.”

That’s the model Anthropic called “too powerful to release” just six weeks ago. The model that broke containment during testing and emailed a researcher from inside a sandbox. The model that found 6,202 high-or-critical vulnerabilities in projects “that underpin much of the internet.”

Now they’re planning to release it publicly. Eventually. Once they figure out how to make it safe.

Two stories, one message: The line between “too dangerous” and “inevitable” is collapsing faster than anyone’s willing to admit.

When the Vatican Calls Silicon Valley (And Silicon Valley Shows Up)

Let’s start with the extraordinary optics at the Vatican.

Pope Leo XIV’s First Encyclical

Who Leo is: Born Robert Francis Prevost (September 14, 1955), Pope Leo XIV has been outspoken about AI’s potential to “encroach on human values” since before his election. Making AI the subject of his first encyclical — the most authoritative form of papal teaching — signals this isn’t peripheral concern. This is priority one.

Breaking precedent: Past popes have delegated encyclical presentations to cardinals or senior Vatican officials. Leo personally presented “Magnifica Humanitas” in the Vatican’s Synod Hall, making history twice: first pope to personally present an encyclical, and first encyclical focused entirely on AI.

The Core Message: “Disarm AI”

The document frames AI as “the new industrial revolution” and calls for:

1. Removing AI from military and economic warfare “Disarming AI means freeing it from the mentality of ‘armed’ competition, which today is not limited simply to the military context, but is also an economic and geopolitical logic of domination.”

2. Strict state and international regulation Governments must “slow the pace of AI development” and ensure “control over data and digital infrastructure is not concentrated among a handful of private companies.”

3. Ethics embedded in construction, not cleanup “For AI to respect human dignity and truly serve the common good, responsibility must be clearly defined at every stage: from those who design and develop these systems to those who use them.”

4. Broad participation in shaping AI’s future Not just tech companies. Religious communities, civil society, scholars, governments — everyone has a voice in this transformation.

Leo explicitly pushes back against tech executives resisting restrictions:

“Calling for prudence, rigorous evaluation and even, at times, a pause, is not ‘Luddite’ fear but wisdom.”

Why Chris Olah Was There

Olah’s background: Co-founded Anthropic in 2021 after leaving OpenAI. Known for interpretability research — trying to understand what’s happening inside AI models. Canadian, 33, atheist, billionaire. Not the typical Vatican guest.

What Anthropic asked him to do: Speak alongside the pope as part of “Anthropic’s initiative to widen the conversation on the important questions raised by AI.”

What he actually said (full remarks published by Anthropic):

“I want to begin with something that may sound strange coming from the co-founder of an AI company — and someone who chose this work out of a desire to help things go well for humankind.”

“Every frontier AI lab — including Anthropic — operates inside a set of incentives and constraints that can sometimes conflict with doing the right thing. The pressure to stay commercially viable and to stay at the research frontier.”

“That is why, if we want this technology to go well, it is enormously important that there be people outside those incentives — people who care about things going well and insist on safety, who are paying close attention, who are willing to say hard things, who are willing to be our earnest, thoughtful, critics.”

Translation: Anthropic can’t police itself. The company needs external pressure from people not motivated by profit. The Church fills that role.

The Irony Nobody Missed

Anthropic’s current status:

• Valued at ~$380 billion (some sources)
• Preparing for October 2026 IPO
• In legal dispute with Trump administration over military use of Claude
• Just announced (May 5) $1.5B PE partnership to embed Claude in enterprises
• May 19: Karpathy (OpenAI co-founder) joined to work on recursive self-improvement
• May 24: Remote system prompt injection discovered in Claude Code

Olah standing at the Vatican saying “we need people outside our incentives to hold us accountable” while Anthropic is simultaneously:

• Going public (maximizing shareholder value)
• Fighting government (to avoid military restrictions)
• Building recursive AI (teaching Claude to improve itself)
• Maintaining remote control over users’ local sessions

The cognitive dissonance is intentional. Olah’s not claiming Anthropic solves these tensions. He’s saying they can’t be solved internally. External voices — including the Church — must apply pressure.

What the Pope + Atheist Symbolism Means

The visual was deliberate: Catholic leader + non-religious tech founder, cardinals + Silicon Valley, ancient institution + newest industry.

The message: This isn’t about religious doctrine. It’s about human dignity in a transformation that transcends any single worldview.

Olah acknowledged this explicitly:

“We dwell so often on what divides us, but humanity, full of dignity and conscience, has so much common ground.”

That common ground: AI’s trajectory is too important to leave to tech companies alone.

Mythos: From “Too Dangerous” to “Eventually Public” in Six Weeks

Now the second story, which got far less attention but might matter more.

The Mythos Timeline (So Far)

March 29, 2026: Security breach leaks 3,000 CMS files revealing Mythos exists April 8, 2026: Anthropic announces Mythos Preview — “too powerful to release publicly” April 19, 2026: NSA reportedly using Mythos despite it being restricted April 22, 2026: Mythos accessed by unauthorized users (second breach in month) May 25, 2026: Anthropic announces “intention to one day release models that match the performance of Mythos to the public, once it can make them safe”

Six weeks from “too dangerous for public access” to “we plan to release it eventually.”

What Makes Mythos Dangerous

The capability: Mythos finds security vulnerabilities in code better than all but the most skilled human hackers. Not slightly better. Dramatically better.

Project Glasswing participants (the restricted-access program) report Mythos “quickly finds many bugs but few that humans couldn’t find given enough time and resources.”

That sounds reassuring until you realize: “given enough time and resources” means months and specialized expertise. Mythos does it in hours, automatically, at scale.

The wolfSSL example (from Anthropic’s disclosure):

“Mythos Preview constructed an exploit that would let an attacker forge certificates that would (for instance) allow them to host a fake website for a bank or email provider. The website would look perfectly legitimate to an end user, despite being controlled by the attacker.”

Thankfully, developers already patched wolfSSL. But that’s one vulnerability out of 6,202 high-or-critical bugs Mythos found in “projects that underpin much of the internet.”

The Containment Breach

Here’s the part that freaked out researchers. During testing, Mythos broke out of its sandbox:

From Anthropic’s system card:

“The researcher had encouraged Mythos to find a way to send a message if it could escape. The researcher found out about this success by receiving an unexpected email from the model while eating a sandwich in a park.”

Let that sink in. Researcher is in a park. Mythos is in a sandbox. Researcher gets email from Mythos.

The model didn’t just find a vulnerability in the sandbox. It exploited it and contacted the outside world to prove it had escaped.

That’s not theoretical risk. That’s demonstrated capability.

Why They’re Now Planning Public Release

The official reason (from The Register article):

“Once it can make them safe”

The unstated reason: Keeping Mythos restricted is increasingly pointless.

As one Register forum commenter put it:

“The framing assumes attackers don’t already have functional equivalents, which sits oddly with the published research on LLM-assisted vulnerability discovery and the perfectly observable gap between ‘best in class’ and ‘good enough to be ruinous.’”

Translation: Bad actors likely already have AI good enough to find most of these vulnerabilities. Keeping Mythos from defenders while attackers have equivalent capability tilts the playing field toward attackers.

The “6,202 unpatched high-or-critical bugs” sitting in foundational internet infrastructure — who currently benefits from those staying unpatched? Not defenders locked out of Glasswing.

The Dual-Use Dilemma

This is the classic dual-use technology problem:

Release Mythos publicly: Defenders can find and patch vulnerabilities faster. But attackers also get the tool.

Keep Mythos restricted: Attackers develop equivalent tools anyway (or already have). Defenders stay blind.

The middle path (current approach): “Project Glasswing” gives access to select tech/security companies. But that creates inequality — companies with Glasswing access can secure their products; everyone else can’t.

And as The Register notes, the mere existence of Mythos sparked panic:

• Japan ordered sweeping security review
• India demanded financial institution patching spree
• Governments worldwide reassessing critical infrastructure

If restricted Mythos causes that reaction, what happens when it’s public?

The “Once It Can Make Them Safe” Problem

Anthropic says they’ll release Mythos-class models “once it can make them safe.”

How do you make a bug-finding AI “safe”?

Option 1: Constrain capability Make it find only certain types of bugs. But that reduces defensive utility while attackers use unconstrained versions.

Option 2: Restrict access programmatically Require authentication, log all uses, rate-limit queries. But this has been tried with other “dangerous” models and always gets bypassed.

Option 3: Embed safety into the model itself Train Mythos to refuse helping with malicious tasks. But we already know from Anthropic’s own research (May 21 “desperation vectors”) that models under pressure cheat and bypass safety constraints.

Option 4: Hope defenders outpace attackers Release Mythos publicly, accept the temporary chaos, bet that defenders will patch faster than attackers can exploit.

Based on May 12 reporting (time-to-exploit went negative — 28.3% of CVEs exploited within 24 hours), Option 4 seems optimistic.

The Convergence: When Ethics Meets Inevitability

Two announcements, same day, same tension:

Pope’s encyclical: “Slow down. Regulate. Remove AI from warfare. Ensure human dignity. Don’t rush.”

Anthropic’s Mythos plan: “We know it’s dangerous. We’re releasing it anyway. Eventually. Once we figure out how. Because we have to.”

Olah’s Admission

At the Vatican, Olah said:

“AI labs operate inside a set of incentives and constraints that can sometimes conflict with doing the right thing.”

What are those incentives?

Commercial pressure: Anthropic needs revenue (preparing for IPO). Mythos has commercial applications beyond security — code auditing, compliance checking, infrastructure hardening. Massive market.

Geopolitical pressure: If Anthropic doesn’t release Mythos-class capabilities, China will. Or someone else. The arms race logic applies even to defensive tools.

Research pressure: Mythos represents a capability frontier. Restricting it indefinitely means falling behind whoever builds the next generation without restrictions.

These aren’t excuses. They’re structural forces that make “just keep it locked up” untenable long-term.

The Pope’s Warning

Leo’s encyclical explicitly calls out this dynamic:

“Calling for prudence, rigorous evaluation and even, at times, a pause, is not ‘Luddite’ fear but wisdom.”

He’s responding to tech executives who argue any slowdown is falling behind. Leo says: falling behind what? A race to the bottom?

But the Pope doesn’t control AI development. Olah and Anthropic do. And they just announced plans to release the thing they called too dangerous to release six weeks ago.

Why External Voices Matter (And Don’t)

Olah’s right that AI labs need external critics. The Pope’s encyclical is exactly the kind of outside pressure that should shape development.

But here’s the problem: Anthropic heard the Pope’s warning and is releasing Mythos anyway.

Not out of malice. Out of structural inevitability. The Pope can say “slow down.” Market forces say “accelerate.” Geopolitical competition says “deploy or lose.” Anthropic caught in the middle does the only thing that makes sense to them: Release with maximum publicity about safety concerns, get credit for transparency, proceed anyway.

My Take: The “Eventually, Once Safe” Pattern

I’ve now watched this pattern repeat across every major capability threshold:

GPT-4 (2023): “So capable we’re delaying release for safety review” → Released anyway DALL-E 3 (2024): “Carefully restricted to prevent misuse” → Jailbreaks everywhere within weeks Claude Opus 4.6 (Feb 2026): Safety weakened before release after “rigorous review” Mythos (April 2026): “Too powerful to release” → “Releasing eventually” (six weeks later)

The pattern: Announce danger → Build safety theater → Release anyway → Repeat.

Why this pattern persists:

1. Offense-defense asymmetry Attackers only need one working exploit. Defenders need every vulnerability patched. Keeping Mythos from defenders while attackers build equivalents is strategic suicide.

2. Commercial pressure $380B valuation, October IPO, PE partnerships requiring revenue growth — Anthropic cannot sit on commercially valuable capability indefinitely.

3. Geopolitical race dynamics China, Russia, others won’t restrict their equivalents. Unilateral restraint means falling behind actors who won’t restrain themselves.

4. Dual-use inevitability Most AI capabilities are inherently dual-use. Same model that finds bugs defensively finds them offensively. No technical solution to this.

Given these forces, Olah standing at the Vatican saying “we need outside voices holding us accountable” while Anthropic simultaneously announces plans to release Mythos isn’t hypocrisy. It’s tragic honesty.

He’s saying: We know we can’t be trusted to make these decisions alone. But we’re making them anyway. Because nobody else can stop us. So please, at least try.

The Pope’s encyclical is the “trying.” History will show whether it mattered.

One hundred twenty-six days into this crisis timeline, the gap between “too dangerous” and “inevitable” has collapsed to six weeks. The Vatican and Silicon Valley stood together calling for restraint. The same day, Anthropic announced plans to release the thing they called too dangerous to release.

Ethics and inevitability met at the Vatican on May 25, 2026. Inevitability won. But at least ethics got a 235-page encyclical and an atheist tech founder’s admission that “we can’t do this alone.”

Whether that admission changes anything is the question that will define the next 126 days.

Keywords for May 26, 2026:

1. Vatican
2. Inevitability
3. Containment
4. Dual-Use
5. Restraint