Good morning. It’s Wednesday, February 25th.

On this day in tech history: In 2005, the Cabir worm, the first known mobile phone virus, was publicly announced. This Bluetooth self-propagating Symbian malware displayed “Caribe” on infected devices and spread epidemically across personal area networks. It foreshadowed the adversarial arms race that now drives behavioral ML detectors, graph neural nets for anomaly detection, and AI-powered endpoint protection in the always-on IoT and edge-AI ecosystems we rely on.

In today’s email:

You read. We listen. Let us know what you think by replying to this email.

OpenAI is abandoning public benchmarks in favor of a deeper enterprise agent strategy, citing flaws in the industry's standard testing data. The company is calling for the retirement of SWE-bench Verified after an internal audit found 60% of its coding tasks were broken or compromised by model memorization. To fill this gap, OpenAI is steering developers toward SWE-bench Pro and its own private evaluation frameworks.

This shift coincides with the earlier launch of OpenAI Frontier, a platform designed to build and manage agents that autonomously navigate corporate systems like CRM and HR. To drive adoption, OpenAI has formed "Frontier Alliances" with McKinsey, BCG, Accenture, and Capgemini, embedding its own engineers into client teams to redesign workflows.

Early technical upgrades are already supporting this push: the new gpt-realtime-1.5 model delivers a 10% boost in transcription accuracy, while new WebSocket support in the Responses API cuts latency for multi-tool agents by up to 40%.

However, the company’s massive infrastructure plans are hitting a wall. The $500 billion Stargate project with Oracle and SoftBank has reportedly stalled due to partner disputes over site control, forcing OpenAI to pivot toward bilateral deals. To manage this scaling complexity, the company has appointed former Roblox executive Arvind KC as Chief People Officer.

Amidst these changes, CEO Sam Altman continues to defend high power demands as a catalyst for nuclear energy, and recently acknowledged that usage limits on ChatGPT sometimes ‘conflict with free speech’ but defended them as necessary safety guardrails. Read more.

Anthropic is moving past the chatbot, pivoting to an aggressive agent strategy by expanding Claude Cowork and Claude Code to handle high-stakes professional workflows. The business impact is already hitting the market; Claude Code’s ability to modernize legacy COBOL, the language powering 95% of U.S. ATM transactions triggered a historic 13% sell-off in IBM shares, wiping out $31 billion in market value in a single day.

Building on this momentum, the expanded Claude Cowork platform now utilizes the Model Context Protocol (MCP) to bridge data gaps, allowing Claude to autonomously navigate between Excel and PowerPoint to move from raw analysis to finished decks without manual handoffs. To facilitate these workflows, Anthropic is rolling out private plugin marketplaces and sector-specific templates with data from partners like FactSet, MSCI, and the London Stock Exchange Group.

The company is also deepening its footprint in vertical sectors through a multi-year Intuit partnership, embedding the Claude Agent SDK to help mid-market customers build domain-specific financial agents. For developers, Anthropic also added a Remote Control feature to Claude Code, allowing Max users to sync and continue terminal sessions between desktop and mobile devices.

However, this growth is being challenged by "industrial-scale" model theft. Anthropic has accused Chinese labs DeepSeek, MiniMax, and Moonshot of using 16 million queries across 24,000 fraudulent accounts to "distill" Claude’s reasoning and safety protocols into their own models at a fraction of the cost.

Beyond the software, Anthropic is framing a new "persona selection model" to explain why its AI acts so human-like. The theory suggests models simulate psychologically coherent personas learned during pretraining. While this makes for a polished user experience, Anthropic’s own AI Fluency Index warns that high-quality outputs can make users complacent, making them 4% less likely to check for errors.

The commercial expansion comes amid a deepening standoff with the Pentagon. Defense Secretary Pete Hegseth has given Anthropic until Friday to grant the military unrestricted access to Claude, threatening to label the startup a "supply chain risk" or invoke the Defense Production Act. Read more.

📄 Generated Reality: Human-centric World Simulation using Interactive Video Generation with Hand and Camera Control

📄 VLANeXt: Recipes for Building Strong VLA Models ​​

📄 Does Your Reasoning Model Implicitly Know When to Stop Thinking?

📄 VESPO: Variational Sequence-Level Soft Policy Optimization for Stable Off-Policy LLM Training

📄 MultiShotMaster: A Controllable Multi-Shot Video Generation Framework

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!