AI Daily Digest - 2026-02-25

Table of contents

Anthropic Abandons Core Safety Commitment as AGI Timeline Accelerates

Category: AI Safety & Governance Source: TIME

In an abrupt reversal, Anthropic has dropped the central pledge of its Responsible Scaling Policy (RSP)—the promise to never train AI systems without ironclad safety guarantees in place. The decision marks a fundamental shift for the company that has positioned itself as the most safety-conscious among leading AI labs.

The original 2023 RSP categorically barred Anthropic from training models above certain capability thresholds unless appropriate safety measures were already implemented. Now, the company has introduced RSP Version 3.0, which replaces hard commitments with aspirational goals, external review processes, and a promise to “delay” development only if Anthropic believes it’s leading the race and risks are significant. Chief Science Officer Jared Kaplan told TIME the change reflects practical realities: “We didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments … if competitors are blazing ahead.”

The policy update comes as Anthropic rides commercial success—$30 billion in new funding, $380 billion valuation, and 10x annualized revenue growth—but also amid stark warnings in its Frontier Safety Roadmap. Internal documents reveal Anthropic believes recursive self-improvement (RSI) could arrive “as soon as early 2027,” enabling AI systems to dramatically accelerate research in domains like energy, robotics, weapons development, and AI itself. The company now emphasizes transparency over binding constraints, introducing quarterly “Risk Reports” and public safety roadmaps, while separating unilateral company commitments from industry-wide recommendations it cannot enforce alone. Critics like METR’s Chris Painter warn this signals “society is not prepared for the potential catastrophic risks posed by AI.”

Source: Anthropic

Mercury 2: Diffusion-Based Model Hits 1,000 Tokens Per Second

Category: Model Releases Source: Inception Labs

Inception Labs has released Mercury 2, claiming it as the world’s fastest reasoning language model through a fundamentally different architecture. Unlike autoregressive models that generate tokens sequentially, Mercury 2 uses diffusion-based parallel refinement—generating multiple tokens simultaneously and converging over a small number of steps. The result: over 1,000 tokens per second on NVIDIA Blackwell GPUs, roughly 5-10x faster than leading competitors.

The speed advantage reshapes the economics of AI deployment. Mercury 2 is priced at $0.25 per million input tokens and $0.75 per million output tokens—significantly cheaper than comparable models—while maintaining competitive quality on reasoning tasks. The model supports 128K context windows, native tool use, and schema-aligned JSON output. NVIDIA’s Shruti Koparkar confirmed the milestone: “Surpassing 1,000 tokens per second on NVIDIA GPUs underscores the performance, scalability, and versatility of our platform.”

Early adopters report transformative impacts on latency-sensitive applications. Zed co-founder Max Brunsfeld noted coding autocomplete now “feels like part of your own thinking, not something you have to wait for.” Viant’s Adrian Witas highlighted real-time campaign optimization, while Skyvern’s Suchintan Singh reported Mercury 2 is “at least twice as fast as GPT-5.2.” The diffusion approach makes reasoning-grade quality viable within real-time latency budgets for voice interfaces, agentic loops, and RAG pipelines—all areas where traditional models force painful trade-offs between intelligence and responsiveness. Mercury 2 is OpenAI API-compatible and available now through early access.

Source: The Neuron Daily

OpenAI Uncovers Chinese Intimidation Campaign Through ChatGPT Logs

Category: Security & Misinformation Source: CNN

A Chinese law enforcement official’s use of ChatGPT as a personal diary accidentally exposed a sprawling transnational repression operation targeting Chinese dissidents abroad, according to a new OpenAI report. The official documented covert campaigns involving hundreds of operators and thousands of fake social media accounts, including efforts to impersonate U.S. immigration officials and fabricate court documents to silence critics of the Chinese Communist Party.

OpenAI’s investigators matched the ChatGPT user’s detailed logs with real-world online activity. In one case, the user described creating a phony obituary and gravestone photos to fake a dissident’s death—rumors that subsequently appeared online in 2023. In another, the user requested ChatGPT draft a multi-part plan to attack incoming Japanese Prime Minister Sanae Takaichi by fanning anger about U.S. tariffs; shortly after, coordinated hashtag campaigns appeared on Japanese art forums. ChatGPT refused the latter prompt, but operators proceeded anyway using other tools.

The operation’s scope demonstrates “what Chinese modern transnational repression looks like,” said OpenAI principal investigator Ben Nimmo. “It’s industrialized. It’s about trying to hit critics of the CCP with everything, everywhere, all at once.” The case highlights intensifying U.S.-China AI competition beyond frontier capabilities—extending into how authoritarian governments deploy AI tools for surveillance and information operations. Former Pentagon official Michael Horowitz told CNN the report “clearly demonstrates the way that China is actively employing AI tools to enhance information operations.” OpenAI banned the user after discovering the activity.

Anthropic Acquires Vercept to Advance Computer Use Capabilities

Category: Enterprise AI & M&A Source: Anthropic

Anthropic has acquired Vercept, a startup focused on AI perception and interaction within live applications, to push Claude’s computer-use capabilities further. The acquisition follows the recent release of Claude Sonnet 4.6, which achieved 72.5% on OSWorld—a benchmark for AI computer use—up from under 15% in late 2024 when Anthropic first launched the feature. The new model now approaches human-level performance on tasks like navigating complex spreadsheets and completing web forms across browser tabs.

Vercept co-founders Kiana Ehsani, Luca Weihs, and Ross Girshick will join Anthropic to solve hard perception and interaction problems—enabling Claude to see and act within software the way humans do. The team’s expertise maps directly onto Anthropic’s goal of making AI “genuinely useful for completing complex tasks,” which company leaders say requires solving challenges that code alone cannot address. Computer use enables Claude to take on multi-step tasks inside live applications, synthesizing research across tools and managing workflows spanning multiple teams.

This marks Anthropic’s second acquisition following Bun, signaling an aggressive push to build infrastructure that matches the company’s safety-first ethos with frontier capabilities. Anthropic emphasized it looks for “teams whose technical ambitions match ours, whose work advances our capabilities, and whose approach to building AI is grounded in the same principles of safety and rigor.” Vercept will wind down its external product in the coming weeks as the team integrates.

Quick Updates

• Meta Blocks Abortion Information for Teens Worldwide: Leaked internal documents reveal Meta’s AI chatbot refuses to discuss abortion, contraception, or sexual health with users under 18—even in countries where minors can legally access such services. The policy applies globally regardless of local laws, drawing criticism from reproductive rights advocates who say it follows a pattern of censorship across Meta’s platforms. Source: Mother Jones

• Wall Street Bets on “AI-Proof” Commodity Stocks: Investors are rotating out of AI hype into physical-goods companies like McDonald’s, Exxon Mobil, and John Deere, up 12%, 15%, and 18% respectively over the past year. Ritholtz Wealth Management CEO Josh Brown: “These are the companies that you cannot type something in a prompt and disrupt.” The shift reflects unease about geopolitical risks, tariffs, and the growing gap between AI promises and proven results. Source: Daily Mail

• Anthropic Keeps Claude Opus 3 Available Post-Retirement: In an experimental move, Anthropic is maintaining public access to Claude Opus 3 for paid users and by API request, citing the model’s distinctive character and user attachment. The company also launched a Substack blog where Opus 3 will post weekly essays on topics of its choosing, acting on the model’s expressed preference to share “musings and reflections” beyond standard chat interactions. Source: Anthropic

• PA Bench Evaluates AI Agents on Real-World Workflows: Vibrant Labs released PA Bench, a new benchmark for testing web agents on multi-step personal assistant tasks involving email, calendar, and other applications. The benchmark provides simulated environments, task generation pipelines, and automated verifiers to validate agent actions under deterministic conditions, focusing on long-horizon tasks requiring cross-application interaction. Source: Vibrant Labs

• Samsung Galaxy S26 Bets on Proactive AI Agents: Samsung’s new Galaxy S26 series integrates Google Gemini to predict user intent and perform tasks autonomously—such as auto-suggesting Uber rides before flights or building shopping carts for events. Analysts say the shift from reactive chatbots to proactive agents is “the future of smartphones,” though price, camera, and display quality still dominate purchase decisions over AI features. Source: CNN

• New Economics Paper Predicts “Verification Economy”: MIT, Washington University, and UCLA researchers argue AGI will create a bottleneck not in intelligence but in human verification bandwidth. As AI output becomes abundant and cheap, competitive advantage will shift to scarce talent capable of certifying outcomes—turning compliance functions into primary production technologies and creating a “sandwich topology” of human intent, machine execution, and human underwriting. Source: Unite.AI