Anthropic's Claude Mythos: A 'Step Change' Revealed by Accident
The biggest AI story of the weekend was an embarrassing security blunder at Anthropic. Security researchers from LayerX Security and the University of Cambridge discovered that a misconfigured content management system had left nearly 3,000 unpublished assets publicly accessible — including a draft blog post describing Anthropic's next-generation model, Claude Mythos.
According to the leaked materials, Mythos represents a new tier of AI models internally called Capybara — larger and more capable than the current Opus lineup. The draft blog stated that compared to Claude Opus 4.6, the new model achieves "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity."
Perhaps most concerning: leaked internal documents warn that Mythos could significantly heighten cybersecurity risks by rapidly finding and exploiting software vulnerabilities. Anthropic also revealed that a Chinese state-sponsored group had used Claude Code in a coordinated campaign to infiltrate approximately 30 organizations before being detected and shut down.
The irony of an AI company building unprecedented cybersecurity capabilities while leaving its own announcements in an unsecured data store was not lost on observers. Anthropic said it is "being deliberate" about the model's release timeline. No public benchmarks have been verified — all performance claims come from the leaked internal descriptions.
Source: Fortune, Fortune (Cybersecurity)
Google Gemini 3 Deep Think Goes Live
Google launched Gemini 3 Deep Think in the Gemini app for Ultra subscribers this weekend, with early API access opening for researchers, engineers, and enterprises. Unlike standard chat models, Deep Think is specifically positioned for harder technical use cases — scientific reasoning, engineering problems, and complex analytical tasks.
This release comes on the heels of an extraordinary month for Google. Gemini 3.1 Pro already dominates 13 out of 16 major performance benchmarks, and the company is simultaneously expanding its Pentagon work through "Agent Designer," a new feature on Gemini for Government that lets military employees build custom AI agents for unclassified tasks.
Google is also deepening its partnership with Apple, providing its 1.2 trillion parameter Gemini model to power a completely reimagined, AI-native version of Siri expected to debut later this year.
Source: The Neuron
OpenAI Quietly Kills the Sora Public API
In one of March's most surprising moves, OpenAI announced the discontinuation of its Sora public API on March 24th, giving developers 30 days' notice. The reason? The economics of video generation at scale simply don't work yet.
The shutdown has already redirected enterprise video AI budgets toward competitors including Runway Gen-4, Pika 2.1, and Google's Veo 2. It's a rare admission from OpenAI that not every frontier AI capability is ready for mass-market deployment, even as the company surpasses $25 billion in annualized revenue.
OpenAI isn't slowing down elsewhere, though. GPT-5.4, released earlier this month in Standard, Thinking, and Pro variants, supports context windows up to 1.05 million tokens and reduces factual claim errors by 33% compared to GPT-5.2.
Source: Digital Applied
White House Releases National AI Policy Framework
On March 20th, the White House released its National Policy Framework for Artificial Intelligence — the most comprehensive federal AI regulatory blueprint to date. The framework calls for federal preemption of state AI laws, argues against creating a new federal AI regulator, and pushes for limited developer liability.
The framework also includes child safety protections, requiring age-assurance for AI platforms accessible to minors and limits on behavioral advertising data collection. On the contentious copyright front, the administration stated its belief that training on copyrighted material does not violate copyright laws, while recommending courts resolve specific disputes.
The EU is moving in parallel: the EU Council agreed on March 13th to streamline certain AI Act rules, extending timelines for high-risk AI system compliance and adding new prohibitions on AI-generated non-consensual intimate content.
Source: CNBC, EU Council
NASA's Mars Rover Makes History with AI-Planned Drives
In a remarkable real-world AI deployment, NASA's Perseverance rover completed the first Mars drives ever planned by artificial intelligence. Using Anthropic's Claude vision-language models, the rover autonomously generated safe waypoints across 456 meters of Martian terrain — replacing a planning task that human operators had performed manually for 28 years.
The achievement represents a significant milestone for autonomous space exploration and demonstrates how frontier AI models are finding applications far beyond chatbots and code generation.
Source: The Neuron