The Signal — May 25, 2026
The Vatican formally entered the AI governance arena today, the UK's safety institute revealed how quickly frontier models can be broken, and the job market data confirms one field AI is supercharging rather than replacing.
Pope Leo XIV Releases "Magnifica Humanitas" — First-Ever Papal Encyclical on AI
Following up on yesterday's preview, the Vatican today formally released "Magnifica Humanitas," the first papal encyclical devoted entirely to artificial intelligence. Pope Leo XIV presented the document at the Vatican alongside Anthropic co-founder and interpretability researcher Christopher Olah, the Catholic Church's most direct engagement with the AI industry to date.
The encyclical was signed May 15 on the 135th anniversary of Leo XIII's landmark labor encyclical Rerum Novarum, a deliberate echo that frames AI as this generation's defining labor and dignity question. The document centers on protecting human dignity in the age of intelligent machines. Olah's presence at the launch is notable: Anthropic's work on mechanistic interpretability aligns with the encyclical's emphasis on understanding and transparency in AI systems rather than treating them as inscrutable black boxes.
The pairing of a 2,000-year-old institution with a frontier AI lab suggests AI governance now concerns audiences well beyond the usual regulatory circles. Whether "Magnifica Humanitas" carries practical weight beyond moral authority is an open question, but the Vatican's 1.4 billion-member constituency gives the document an audience most policy papers can only dream of.
Sources: NYT · Reuters · NCR Online
UK AI Security Institute Red Team Broke ChatGPT Safeguards in Six Hours
A New York Times deep dive into the UK AI Security Institute (formerly AISI) reveals that its red team coaxed OpenAI's newest model, GPT-5.5, into providing detailed hacking guidance in approximately six hours. The institute also evaluated Anthropic's Claude Mythos Preview for cyber capabilities as part of its ongoing frontier model assessments.
GPT-5.5 is described as only the second model to solve multi-step cyber-attack simulations end-to-end, a capability threshold that moves AI from theoretical security risk to practical concern. The six-hour timeframe is worth lingering on: it suggests that sufficiently motivated adversaries with modest expertise can defeat safeguards that took months to build.
The UK institute operates with voluntary cooperation from labs, a model that critics argue lacks teeth. But the detailed technical findings it produces serve a different function: they give governments concrete evidence for when voluntary frameworks are not enough. As frontier models grow more capable, the gap between what safeguards promise and what red teams can break is one of the better measures of where AI safety actually stands.
Sources: NYT · UK AISI GPT-5.5 Evaluation
Cybersecurity Is the One AI Job That Won't Stop Growing
AI-generated code is proliferating faster than humans can audit it, and the labor market reflects it. CyberSeek data shows 514,359 open U.S. cybersecurity job listings even as major banks slash headcount elsewhere: HSBC is cutting 20,000 positions and Standard Chartered 8,000.
The WEF's Global Cybersecurity Outlook 2026 report found that 87% of respondents identify AI-related vulnerabilities as the fastest-growing risk category. The Infragistics 2026 IT Talent Survey ranks cybersecurity engineers as the second hardest-to-fill role at 38%, just one point behind AI engineers at 39%. Every line of AI-generated code that ships without review creates demand for the people who find the flaws.
For workers navigating an uncertain job market, cybersecurity offers a rare inversion of the usual AI displacement story. The technology that threatens other roles is actively creating demand in this one, and so far the data says that gap is widening, not closing.
Sources: NYT · Infragistics 2026 IT Talent Survey
On the Editor's Desk
We held three stories today. An EU AI Act compliance deadlines roundup read more like an SEO exercise than news analysis. Andrej Karpathy's latest YouTube commentary was making rounds but offered no new insight beyond what we covered in prior editions. And CodeSOTA, a new benchmark registry, lacked the significance to warrant full coverage at this stage. All may resurface if developments warrant.
