#1Anthropic's "Claude Mythos" Leaked — Then Confirmed
A misconfiguration in Anthropic's CMS exposed nearly 3,000 internal documents, revealing an unreleased model called "Claude Mythos." After Fortune broke the story, Anthropic confirmed it's already testing Mythos with select customers, calling it a "step change" in reasoning, coding, and cybersecurity capabilities — though also flagging "unprecedented cybersecurity risks."
Source: [Fortune](https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/) | [The Decoder](https://the-decoder.com/anthropic-leak-reveals-new-model-claude-mythos-with-dramatically-higher-scores-on-tests-than-any-previous-model/)
---
#2OpenAI's Next Model "Spud" Has Finished Pretraining
OpenAI is quietly preparing a major release: a new model codenamed "Spud" has completed pretraining, likely positioning for a post-GPT-5.4 rollout. Meanwhile, OpenAI has crossed $25B in annualized revenue and is taking early steps toward a public listing.
Source: [NeuralBuddies](https://www.neuralbuddies.com/p/ai-news-recap-march-27-2026) | [Crescendo AI](https://www.crescendo.ai/news/latest-ai-news-and-updates)
---
#3Google Gemini 3.1 Flash-Lite Drops at $0.25/M Tokens
Google pushed Gemini 3.1 Flash-Lite into general availability — 2.5× faster than prior versions and priced aggressively at $0.25 per million input tokens. Its sibling, Gemini 3.1 Flash Live, is in developer preview for real-time agent use cases.
Source: [Promethean AI Week in Review](https://www.promethean-ai.com/post/the-week-in-ai-news-march-22-27-2026)
---
#4NASA Lets Claude Plan Mars Rover Drives — A First
NASA's Perseverance rover completed its first drives planned entirely by AI, using Anthropic's Claude vision-language models to analyze orbital imagery and terrain data. The AI replaced a manual planning process human operators had performed for 28 years.
Source: [Crescendo AI](https://www.crescendo.ai/news/latest-ai-news-and-updates)
---
#5Three Charged for Smuggling $2.5B in AI Chips to China
The DOJ charged three individuals — including a cofounder of Super Micro Computer — with conspiring to export billions in NVIDIA GPUs to China in violation of US export controls. The case signals aggressive federal enforcement of AI hardware sanctions.
Source: [Al Jazeera](https://www.aljazeera.com/news/2026/3/20/three-charged-in-the-us-with-smuggling-ai-chips-into-china) | [DOJ](https://www.justice.gov/opa/pr/three-charged-conspiring-unlawfully-divert-cutting-edge-us-artificial-intelligence)
---
#6GPT-5.4 Scores Above Human Experts on Economic Tasks
OpenAI's GPT-5.4 "Thinking" model hit 83.0% on the GDPVal benchmark — above the threshold for human expert performance on economically valuable tasks — and 75% on OSWorld-V for autonomous software execution. The 1M-token context window is now live for paid users.
Source: [devFlokers](https://www.devflokers.com/blog/ai-news-march-24-2026-releases-breakthroughs)
---
#7OpenAI Kills Sora's TikTok App
OpenAI is shutting down the Sora mobile app — a short-form AI video platform launched in France last October — despite the underlying Sora 2 model being widely praised. The move suggests OpenAI is pulling back from consumer social distribution in favor of API and enterprise focus.
Source: [NeuralBuddies](https://www.neuralbuddies.com/p/ai-news-recap-march-27-2026)
---
**Big Picture**
Today's throughline is capability overhang colliding with institutional chaos. Anthropic's Mythos leak, GPT-5.4 clearing human-expert benchmarks, and AI now autonomously navigating Mars all point to a technology moving faster than the organizations building it can manage. The same week that AI is piloting rovers on another planet, a company is accidentally leaking its most powerful unreleased model due to a default file permission setting. The frontier is widening — but so are the cracks in the infrastructure holding it together.
