Monday, May 4, 2026

AI Daily Brief  ·  8 stories  ·  prates.fyi
Anthropic Apr 7, 2026

Anthropic Launches Claude Mythos Preview in Closed Cybersecurity Initiative

Anthropic unveiled Claude Mythos Preview, its most powerful model to date — but it won't be available to the public anytime soon. The model demonstrated an alarming ability to find and exploit zero-day vulnerabilities across every major operating system and web browser. Anthropic is limiting deployment to 12 handpicked partners under Project Glasswing, a defensive security program backed by $100M in usage credits and $4M in open-source security donations.

Amazon Apple Broadcom Cisco CrowdStrike Linux Foundation Microsoft Palo Alto Networks
Mythos can reverse-engineer exploits on closed-source software and turn known (N-day) vulnerabilities into working exploits. Anthropic is withholding general release until protective safeguards are in place.
cybersecurity zero-day limited-preview project-glasswing
TechCrunch
Google Apr 24, 2026

Google Commits Up to $40 Billion in Anthropic, Its Biggest AI Bet Yet

Google announced a landmark investment of up to $40 billion in Anthropic — a combination of cash and cloud compute credits — cementing the search giant's position as Anthropic's largest backer. The deal follows Anthropic's limited release of the Mythos model and signals Google's long-term confidence in Claude as a competitive alternative to OpenAI's frontier models.

$40B Google commitment
cash + compute investment form
investment google-cloud funding
TechCrunch
OpenAI Apr 23–24, 2026

GPT-5.5 Ships with Record Benchmarks; Codex Gets Major Upgrade

OpenAI released GPT-5.5, now powering Codex, its agentic coding platform. The model sets new state-of-the-art scores on coding benchmarks and excels at computer use, long-horizon research, and multi-step task completion. It runs on NVIDIA GB200 NVL72 rack-scale infrastructure, with over 10,000 NVIDIA employees already using the system internally.

82.7% Terminal-Bench 2.0
58.6% SWE-Bench Pro
Available in API + ChatGPT (Plus/Pro/Business/Enterprise). GPT-5.5 Pro variant rolling out to Pro and above tiers.
gpt-5.5 codex agentic-coding swe-bench
openai.com
Google Google Cloud Next '26 · Apr 2026

Google Launches Gemini Enterprise Agent Platform at Cloud Next '26

Google unveiled the Gemini Enterprise Agent Platform, a unified surface to build, govern, scale, and optimize enterprise AI agents. The platform ships with Agent Studio (low-code, for non-technical users), the Agent Development Kit (ADK) (code-first, graph-based reasoning), and deep integration with Google Workspace and Cloud. The underlying Gemini 3.1 Ultra model offers a 2-million-token context window across text, image, audio, and video.

85.9 Gemini 3.1 Pro BrowseComp
2M tokens context window
gemini-3.1 enterprise-agents agent-studio adk
Google Cloud Blog
Industry May 1, 2026

Pentagon Signs AI Deals With Seven Companies — Anthropic Conspicuously Absent

The Department of Defense struck agreements with OpenAI, Google, Microsoft, Amazon, Nvidia, SpaceX/xAI, and Reflection AI to deploy AI within classified Pentagon networks for "lawful operational use." Anthropic was excluded after it refused to grant the DoD unrestricted access to Claude for use in fully autonomous weapons and mass domestic surveillance — a designation the Pentagon formally classified as a supply-chain security risk in March.

A possible reconciliation is on the table: Trump said a deal with Anthropic is "possible" after CEO Dario Amodei met with White House officials in April. Google and OpenAI included language in their contracts limiting autonomous weapons and mass surveillance use.
defense pentagon policy anthropic classified-networks
Washington Post
Microsoft May 1, 2026

Microsoft Agent 365 Goes GA; M365 E7 "Frontier Suite" Launches

Microsoft made Agent 365 generally available alongside Microsoft 365 E7, the new "Frontier Suite" tier. Agent 365 extends Entra network controls to Copilot Studio agents running on endpoint devices, introduces a Shadow AI discovery page for visibility into unsanctioned local agents, and lets IT admins manage AI tools like OpenClaw and Claude Code via Intune policies — a significant governance milestone for enterprise AI.

Accenture is deploying Copilot to 743,000 workers in what Microsoft calls the biggest enterprise AI productivity test at scale to date.
agent-365 m365-e7 shadow-ai enterprise-governance
Microsoft Security Blog
xAI Feb 2, 2026 · ongoing

SpaceX Acquires xAI in $1.25T Merger; Grok 5 Expected Q2

SpaceX completed its acquisition of xAI in February, creating the largest merger ever at a combined $1.25 trillion valuation. Grok is now a wholly owned subsidiary of SpaceX, with plans to build space-based AI data centers to address global power constraints. Meanwhile, Grok Imagine 1.0 launched with 10-second 720p video generation. Grok 5, with 6 trillion parameters and native video understanding, is expected in Q2 2026.

$1.25T combined valuation
6T params Grok 5 (Q2 2026)
spacex grok-imagine video-gen grok-5 merger
TechCrunch
Cursor Apr 2, 2026

Cursor Composer 2 Ships on Kimi K2.5 with Top-Tier SWE-Bench Scores

Cursor released Composer 2, built on the Kimi K2.5 base model, delivering strong performance on multilingual code repair and terminal-task benchmarks. The release comes as Claude Code emerged as the most-used AI coding tool in a February 2026 Pragmatic Engineer survey of 906 developers, with Cursor and Copilot trailing. Both tools now converge on background agents, CLI access, and autonomous task completion.

73.7% SWE-bench Multilingual
61.7% Terminal-Bench 2.0
cursor composer-2 kimi-k2.5 swe-bench agentic-coding
The New Stack