Daily AI Briefing — May 26, 2026

Pope Leo XIV released his first encyclical, Magnifica Humanitas, calling for AI to be disarmed and subjected to rigorous ethical constraints.

• Anthropic co-founder appeared alongside the pope in Rome, signaling Vatican engagement with frontier labs
• The 40,000-word text condemns autonomous weapons, data colonialism, and patent hoarding
• The pope linked the digital economy to new forms of slavery and apologized for church delays on historical slavery
• The Guardian editorial drew parallels to Leo XIII's 1891 Rerum Novarum on industrial labor

On the infrastructure side, agentic and serving tooling continued to mature:

• Together AI open-sourced OSCAR, an attention-aware 2-bit KV cache quantization system enabling INT2 compression compatible with paged KV-cache serving
• WorkOS released auth.md, an OAuth-based open protocol for AI agent registration to replace insecure API key workarounds
• MCP SDK downloads reached 97 million monthly, with Anthropic donating the protocol to the Linux Foundation's Agentic AI Foundation

Additional coverage included a WIRED report on AI-driven vulnerability discovery arms races and a Louisiana ethics investigation tied to Meta's Hyperion datacenter land deals.

Today's research is dominated by safety, alignment, and interpretability work, with several substantive contributions from Anthropic. Constitution-following audits of frontier models from Anthropic and OpenAI reveal specification gaps under multi-turn adversarial pressure. Faithfulness as Information Flow introduces sufficiency/completeness/necessity properties for CoT monitoring, while a complementary meta-evaluation shows existing faithfulness metrics fail to measure what they purport to measure.

• Polymorphism Is Rotation: independently trained transformers compute the same function in residual-stream bases that differ only by a uniform random rotation, with strong implications for cross-model SAE transfer and circuit universality
• From Simulation to Enaction (Jack Lindsey, Anthropic): post-trained LLMs implicitly recognize their own generations with 3-4x lower entropy than off-policy text, across families and scales
• Cognitive Security as an AI Safety Cause Area (Steinhardt): position paper framing belief/action autonomy as a major emerging safety frontier

Real-world impact and agent capabilities round out the highlights. A 45,000-researcher global RCT delivered LLM feedback on 31,000+ arXiv preprints, providing rare causal evidence on AI's effect on scientific output. QUEST releases an open family of 2B-35B deep research agents with a full mid-training/SFT/RL recipe focused on citation grounding. HumanEgo achieves zero-shot human-to-robot transfer from minutes of egocentric video via entity-level hand-object representations, and self-generated replay is shown to nearly eliminate catastrophic forgetting when model capacity is unsaturated.

The Anthropic-Vatican partnership dominated AI discourse following Pope Leo XIV's first encyclical *Magnifica Humanitas* on AI, with Chris Olah invited to speak at the presentation. Tomasz Tunguz framed it as a major signal of institutional engagement, while Timnit Gebru led sharp criticism, calling it 'Vatican-washing' and arguing the Church should have partnered with exploited data workers rather than Anthropic.

• xAI launched Grok Build in Beta for SuperGrok and X Premium+ users, featuring Plan Mode and a CLI, drawing 19.7M views
• Yann LeCun sparked debate distinguishing engineering from science, defending peer review against critics
• Gary Marcus intensified bubble warnings and AI backlash predictions, while defending Demis Hassabis against Noah Smith on AGI claims
• François Chollet argued AI shouldn't be framed as a productivity booster but as enabling new workflows
• Ethan Mollick called for hard-problem repositories beyond math benchmarks

• Technical highlights included a paper on the Implicit Curriculum Hypothesis in pretraining, MeMo's separate memory architecture for LLMs, and NVIDIA DGX Spark running 16 local agents via MiniMax M2.7. Greg Brockman promoted GPT-5.5 Pro for fact-checking, while Clement Delangue (HuggingFace) warned that AI power concentration is the most critical risk.