🔥 Anthropic Investigates Reports of Unauthorized Access to Mythos AI
Anthropic is investigating claims that a rogue group gained unauthorized access to its advanced Mythos AI model — a system the company itself has described as a cybersecurity “reckoning.” Reports from Bloomberg, The Guardian, and WIRED indicate that U.S. Treasury Secretary Powell and Secretary Bessent have briefed major banks on the situation. The controversy has escalated to the highest levels of government, with Axios reporting Anthropic withheld the Mythos Preview model specifically because its hacking capabilities were deemed too powerful.
🚀 DeepSeek Makes 75% Price Cut Permanent on V4-Pro
DeepSeek has permanently slashed the price of its flagship V4-Pro model by 75%, pricing output tokens at least 34x below GPT-5.5. This aggressive move escalates the ongoing AI price war, making high-performance inference dramatically more affordable. The permanent price cut signals DeepSeek’s confidence in its cost structure and positions the Chinese AI lab as a formidable competitor in the global market.
🤖 Alibaba Launches Qwen3.7-Max: Autonomous AI Agent for 35 Hours
Alibaba has unveiled Qwen3.7-Max, a reasoning agent model with a 1M-token context window that can operate autonomously for 35 hours and execute over 1,000 tool calls in a single session. The model supports external harnesses including Anthropic’s Claude Code, and was demonstrated optimizing code for Alibaba’s own custom chips. Qwen3.7-Max will also be integrated into Taobao for agentic shopping experiences and into automotive systems for voice-controlled services.
🔄 Meta Abandons Open-Source Llama, Launches Proprietary Muse Spark
Meta has released Muse Spark, its first proprietary AI model since the formation of its Superintelligence Labs. The move marks a strategic departure from the open-source Llama family that had defined Meta’s AI strategy for years. Muse Spark is described as a rebuilt AI stack after Llama’s perceived underperformance, and represents Meta’s attempt to monetize its AI investments more directly through proprietary offerings.
💰 OpenAI Prepares Trillion-Dollar IPO
OpenAI is preparing to confidentially file for its initial public offering in the coming weeks, with the IPO potentially valuing the company at over $1 trillion. The move comes as OpenAI has been named a Leader in enterprise coding agents by Gartner and continues to make breakthrough research progress, including solving an 80-year-old mathematics problem.
💸 xAI Burned $6.4 Billion Last Year
According to SpaceX’s S-1 IPO filing, Elon Musk’s xAI burned $6.4 billion in the last fiscal year — approximately $1.3 billion per month — while generating only ~$500M in revenue. The filing reveals that Grok’s spending is heavily subsidized by SpaceX’s operations, raising questions about the sustainability of xAI’s aggressive infrastructure buildout.
🔐 NSA Releases Security Guidance for Model Context Protocol (MCP)
The National Security Agency (NSA) has published security design considerations for AI-driven automation leveraging the Model Context Protocol (MCP). This marks a significant milestone for the MCP ecosystem, as the guidance addresses authentication, authorization, and secure deployment patterns for agentic AI systems — directly relevant to the growing MCP developer community.
🤗 Trending on Hugging Face
Notable new models uploaded in the last 24 hours include:
- iamseungpil/metacot-h200-e20a-repro-sft-0522 — A Qwen3-based fine-tune with reinforcement learning from scratch (19 downloads)
- yorickr/Qwen3.6-27B-MTP-GGUF — GGUF quantized version of Qwen3.6-27B with multi-token prediction support
- mradermacher/Stylizer-V2-LLaMa-70B-heretic-i1-GGUF — A mergekit fusion model combining Stylizer and LLaMa architectures
- lt2c/metamath-qwen3.5-0.6b-rl-gamma1 — Small-scale RL fine-tune using TRL on Qwen3.5-0.6B for math reasoning
- Ryscu/medgemma-4b-neuroOnco — Specialized medical model for neuro-oncology applications
⭐ GitHub Trending: AI Edition
| Repository | Stars | Language | Description |
|---|---|---|---|
| Doorman11991/smallcode | 1,379 | JavaScript | AI coding agent optimized for small LLMs — 87% benchmark with 4B-active model |
| lynote-ai/humanize-text | 649 | Python | Open-source AI text humanizer to bypass AI detectors |
| basketikun/infinite-canvas | 478 | TypeScript | Open-source infinite canvas for AI creative workflows |
⏰ Limited-Time Free Models (Expiring Soon)
| Model | Expires | Pricing (per M tokens) | Context |
|---|---|---|---|
| Baidu: Qianfan-OCR-Fast | May 28 (2 days) | $0.68/M prompt · $2.81/M completion | 65K |
| Mistral: 7B Instruct v0.1 | May 30 (4 days) | $0.11/M prompt · $0.19/M completion | 4K |
| Google: Gemini 2.0 Flash Lite | Jun 1 (6 days) | $0.075/M prompt · $0.30/M completion | 1M |
| Google: Gemini 2.0 Flash | Jun 1 (6 days) | $0.10/M prompt · $0.40/M completion | 1M |
💡 Key Trends
This Sunday’s news cycle underscores several accelerating themes: AI safety is now a geopolitical concern — the Anthropic Mythos incident and NSA MCP guidance both point to governments scrambling to regulate agentic AI. The AI price war is far from over — DeepSeek’s permanent 75% cut and the steady stream of free-tier models expiring suggest margins are being compressed across the industry. Proprietary beats open-source for now — Meta’s shift from Llama to Muse Spark signals a broader industry recalibration around monetization. Meanwhile, the IPO pipeline (OpenAI, SpaceX, Anthropic) suggests the AI sector is maturing into a publicly-traded industry.