Tuesday, May 5, 2026

AI Daily Brief  ·  8 stories  ·  prates.fyi

Today at a Glance

Anthropic

Claude Mythos Finds Thousands of Zero-Days — But Stays Off the Shelf

Anthropic's unreleased frontier model, Claude Mythos Preview, autonomously identified and exploited zero-day vulnerabilities across every major operating system and web browser — including a 17-year-old remote code execution flaw in FreeBSD (CVE-2026-4747). The company has deliberately kept the model out of general availability due to fears it could enable large-scale AI-driven cyberattacks.

Under Project Glasswing, Anthropic is giving select partners — including AWS, Apple, Cisco, Google, JPMorgan Chase, Microsoft, Nvidia, and the Linux Foundation — early access to find and patch vulnerabilities before adversaries do. The initiative is framed as a race to let defenders lead. Separately, Anthropic also launched Claude Security, an enterprise product with more limited capabilities than Mythos.

Sources: anthropic.com · The Hacker News · red.anthropic.com

Defense / Policy

Pentagon Signs AI Deals with 8 Tech Giants — Anthropic Notably Absent

The U.S. Department of Defense has signed AI contracts with OpenAI, Google, Microsoft, Amazon, Oracle, Nvidia, SpaceX/xAI, and Reflection AI — pointedly excluding Anthropic. The exclusion stems from Anthropic's refusal to allow Claude to be used for "all lawful purposes," including autonomous weapons and mass surveillance.

Defense Secretary Pete Hegseth called Anthropic CEO Dario Amodei an "ideological lunatic" during congressional testimony. Anthropic responded by suing the Trump administration; a federal judge in California has since blocked the government's effort to compel compliance. SpaceX qualified as a vendor through its 2026 merger with xAI Holdings, giving Elon Musk's company a DoD foothold for Grok models.

Sources: gHacks · CNN Business

OpenAI

GPT-5.5 Rolls Out to ChatGPT Subscribers; API Coming "Very Soon"

OpenAI announced GPT-5.5, its newest frontier model, now rolling out to Plus, Pro, Business, and Enterprise ChatGPT users and in Codex. The model sharpens multi-step reasoning, tool use, coding, autonomous research, and document creation — less than two months after GPT-5.4 shipped. OpenAI described it as the company's "smartest model yet for professional work."

API availability is imminent, with the company citing additional safety and security reviews for enterprise-scale serving. GPT-5.5 also became available on Amazon Bedrock, alongside OpenAI's Codex and Bedrock Managed Agents, giving AWS enterprise customers access through their existing infrastructure.

Sources: CNBC · openai.com/news

OpenAI

OpenAI Crosses $25B in Annualized Revenue, Explores Late-2026 IPO

OpenAI has surpassed $25 billion in annualized revenue and is reportedly taking early steps toward a public listing, potentially as soon as late 2026. The company also locked in a landmark compute deal with Cerebras — up to 750 MW of capacity valued at more than $20 billion through 2028 — reducing its dependence on scarce NVIDIA H100/H200 inventory.

The milestone underscores a dramatic reversal from the company's nonprofit origins: OpenAI now generates more revenue than most Fortune 500 companies that have been public for decades.

Sources: openai.com/news

Google

Gemini 3.1 Flash-Lite Is Now in Preview: 2.5× Faster, 45% More Output, $0.25/1M Tokens

Google launched Gemini 3.1 Flash-Lite in developer preview via Google AI Studio and Vertex AI. The model delivers a 2.5× improvement in Time-to-First-Token and a 45% increase in output generation speed compared to Gemini 2.5 Flash, while maintaining similar or better quality on key benchmarks — and is priced at just $0.25 per million input tokens / $1.50 per million output tokens.

In a related move, Google's Gemini iOS app received a "Liquid Glass" visual redesign, and eagle-eyed users spotted references to upcoming Gemini 3.2 Flash and Gemini 3.1 Lite in the app's metadata — hinting at announcements likely timed around Google I/O.

Sources: blog.google · Piunika Web

xAI · Microsoft

xAI's Grok 4.1 Fast Joins Microsoft Copilot Studio Alongside OpenAI and Anthropic

Microsoft added xAI's Grok 4.1 Fast to Copilot Studio's model provider lineup, enabling U.S.-based enterprise builders to integrate it into low-code AI workflows. Grok 4.1 Fast is optimized for large-context window tasks, deep tool use, and complex multi-step reasoning, with real-time grounding via X's live data stream.

The move cements Microsoft's multi-model strategy — Copilot Studio now offers OpenAI, Anthropic, and xAI models in one environment — and signals that no single vendor will hold a monopoly inside Microsoft's enterprise stack. xAI also separately released Grok 4.20, targeting factuality on current-events queries with improved source attribution.

Sources: microsoft.com · x.ai/news

xAI

xAI Closes $20B Series E as SpaceX Merger Embeds Grok in Autonomous Spacecraft

xAI completed its upsized Series E funding round at $20 billion, exceeding the original $15 billion target. The raise comes after the formal merger between SpaceX and xAI Holdings, which Elon Musk announced would deeply embed Grok models into SpaceX's operations — targeting fully autonomous spacecraft and, eventually, robotic Mars colony infrastructure.

The SpaceX-xAI combination also qualified the entity for the Pentagon's new AI vendor contracts, giving the merged company a direct U.S. defense foothold that pure AI labs like Anthropic currently lack.

Sources: x.ai/news

AI Coding

Cursor 3 Launches Agent Mode as Codex Tops 3M Weekly Users; Claude Code Leads Sentiment

Cursor launched Cursor 3, a redesigned interface that lets developers spin up autonomous AI coding agents from within the IDE — a direct response to Anthropic's Claude Code and OpenAI's Codex. The three products are evolving around distinct niches: Codex as cloud-based fire-and-forget automation, Cursor as an interactive visual IDE, and Claude Code as a terminal-native assistant for deep codebase reasoning.

Growth metrics underscore the market's momentum: Codex surpassed 3 million weekly active users (up from 2M a month ago). A Pragmatic Engineer survey of 906 engineers found Claude Code ranked #1 in "most loved" at 46%. SemiAnalysis estimates Claude Code now drives roughly 4% of all public GitHub commits, with a projection of 20% by year-end.

Sources: The New Stack · Startup News