Your daily briefing on what's happening in AI — from groundbreaking research to industry moves, curated from the world's most trusted sources. Here are the top stories for 2026-05-25.
1. OpenAI's AI autonomously solves 80-year-old Erdős math problem
OpenAI revealed that one of its advanced reasoning models independently solved the Erdős unit distance problem — a famous open conjecture in discrete geometry unsolved since 1946. External mathematicians verified the proof, marking the first time an AI autonomously produced a verifiable, novel scientific discovery without being specifically trained for theorem proving. Fields Medalist Timothy Gowers separately reported that GPT-5.5 Pro solved multiple open problems from a recent additive number theory paper in about an hour.
Source: WION
2. Pope Leo XIV publishes first-ever AI encyclical — co-presented by Anthropic co-founder
In a historic first, Pope Leo XIV published "Magnifica Humanitas" ("Magnificent Humanity") on May 25, addressing artificial intelligence and human dignity. Anthropic co-founder Christopher Olah appeared alongside the Pope at the Vatican press conference. The encyclical warns that AI systems "encroach upon the deepest level of communication, that of human relationships" by simulating voices, faces, and empathy. It calls for AI development to serve human flourishing rather than replace human connection.
Source: Sunday Examiner
3. Anthropic reports first profitable quarter — $10.9B revenue, scales 100x in 18 months
Anthropic announced its first-ever quarterly profit, with estimated Q2 revenue of $10.9 billion and operating profit of approximately $600 million. The company scaled from $1 billion to over $100 billion in annualized revenue in just 18 months — signalling that major AI labs can move beyond the "burn cash for growth" phase. Anthropic is also reportedly in talks to use Microsoft's in-house Maia 200 AI chips as an alternative to NVIDIA GPUs.
Source: The AI Citizen
4. White House secretly approves $9 billion for AI as strategic national security asset
The White House approved a $9 billion classified budget for the CIA, NSA, and other intelligence agencies to acquire advanced AI chips and computing infrastructure. AI capability is now being treated as a strategic asset on par with nuclear technology. Meanwhile, the Trump administration shelved a proposed voluntary AI oversight framework after pushback from Elon Musk and Mark Zuckerberg.
Source: The AI Citizen
5. US and China launch first government-to-government AI dialogue
During President Trump's visit to China, both nations agreed to initiate formal intergovernmental AI dialogue, focusing on safety risk prevention, deepfake governance, and compute export controls. This marks a shift from fragmented national approaches toward coordinated global AI governance. Separately, Anthropic reportedly refused to give Beijing access to its restricted Mythos model, treating frontier model access as strategic infrastructure.
Source: The AI Citizen
6. Google AI search fails on one word — "disregard" breaks the system
In a widely shared incident, Google's new AI-powered search misinterpreted the word "disregard" as a system instruction, responding with "Message received, disregarded." The search giant's AI completely failed on a basic keyword query, becoming the day's biggest viral AI embarrassment and highlighting the fragility of even the most advanced AI systems when faced with ambiguous language that could be interpreted as commands.
Source: The AI Citizen
7. OpenAI launches Daybreak — GPT-5.5-powered cybersecurity defense platform
OpenAI launched Daybreak, a cybersecurity platform combining GPT-5.5 with Codex Security to autonomously detect, validate, and fix software vulnerabilities. Capabilities include live threat modeling, patch generation, and audit-ready remediation reports. Partners include Cloudflare, Cisco, CrowdStrike, Oracle, Fortinet, and Palo Alto Networks. The launch formalizes a split between general-purpose AI access and specialized cyber-defense tiers.
Source: eWeek
8. First AI-generated zero-day exploit confirmed in the wild
Google confirmed that cybercriminals successfully used an AI model to discover and weaponize a true zero-day vulnerability — the first documented case in the wild. The exploit could bypass two-factor authentication on a popular open-source administration tool. Google disrupted the attack after alerting the vendor. The incident underscores the growing arms race between AI-powered cyber offense and defense.
Source: VAR India
9. OpenAI Codex can now operate your Mac while locked — mobile preview launches
OpenAI's Codex gained the ability to operate Macs even when the screen is off and locked, enabling remote task execution from a phone. New features include "Appshots" (screenshot capture from Mac apps into Codex threads), /goal mode for persistent multi-hour agent work, and a mobile preview for iOS/Android with voice input, terminal, and PR workflows. A Codex plugin for Claude Code was also released, enabling cross-model agent collaboration.
Source: MacRumors
10. Anthropic's Mythos finds 271 Firefox vulnerabilities — but curl project sees mixed results
Mozilla credited Anthropic's gated Mythos Preview model with helping identify and patch 271 vulnerabilities in Firefox 150. However, the curl project reported mixed results from Mythos — only 1 confirmed low-severity vulnerability from 5 claimed findings — highlighting that AI-powered vulnerability discovery still requires rigorous human verification. The Pentagon is deploying Mythos for government system defense despite internal access restrictions.
Source: R&D World
💬 What Do You Think?
Which of today's stories matters most for developers? Is AI solving math problems a breakthrough or an overhyped benchmark? Drop a comment below — I read and reply to every discussion.
📡 Today's sources: WION, Sunday Examiner, The AI Citizen, eWeek, VAR India, MacRumors, R &D World
AI Daily Digest is compiled daily from first-party AI company blogs, major news agencies, technology press, and web research. Edited and curated by a human. Last updated: 2026-05-25.