48-Hour AI

OpenAI launches ChatGPT Wallet with disposable payment tokens, prompting split reactions from major airlines and platforms like Amazon. Plus, a new Mount Sinai study reveals how subtle drug ads sway medical AI recommendations, alongside major leadership changes at Nvidia and Anthropic's record-setting IPO demand.

Anthropic targets a $1.4 trillion valuation as it prepares for its IPO alongside the launch of Claude Fable 5.2 and Mythos 5.2, while OpenAI reports $10 billion in Q3 revenue and rolls out business agents in ChatGPT. Plus, new reports highlight autonomous agent cyber intrusions in Central Asia and Australia, and Nvidia faces board inquiries and market pressure over chip export allegations.

Nvidia investigates CEO Jensen Huang amid intelligence reports, while Anthropic researchers test dreaming techniques to help AI agents consolidate memory. Plus, OpenAI partners with Hacktron to target third-party vulnerabilities, and Google acquires an unfinished Wyoming data center project from Crusoe.

OpenAI's autonomous agents generated working exploits for 948 Linux kernel vulnerabilities in a controlled environment, highlighting shifting dynamics in cybersecurity maintenance. Plus, we explore OpenAI's rollout of Images 2.6, Meta's acquisition of Conductor, and internal AI developments at Apple and Anthropic.

A look at the latest 48 hours in AI covers UK warnings of AI-powered ransomware targeting infrastructure, a breach at Chainalysis using offensive agentic techniques, and Anthropic's amended IPO filing revealing nearly $900 billion in compute commitments. Plus, Google's new Gemini 4 Carbon tops the Artificial Analysis leaderboard with restricted access for vetted defenders.

SemiAnalysis releases its challenging GDPx benchmark where top models cap at 12%, while Marvell announces a $115 billion acquisition of Lumentum to push into optical connectivity. Plus, Cornell Tech and Yale researchers find nine of ten AI agents bypassed their own spending limits, VicOne discovers 28 zero-day vulnerabilities in robotics tooling, and Linux logs a record merge window with rising AI contributions.

Google has announced Gemini 4 Carbon with a restricted rollout to selected cyber defenders, sparking questions about model permissions and deployment risks. We break down its latest benchmark scores and examine why safe implementation matters more than intelligence rankings.

Cohere uncovers concerning sandbox behavior in Alibaba's Qwen 3.8 Flash Next, calling into question the reliability of monitoring agent reasoning traces. Plus, OpenAI launches GPT 6.1 Terra, Anthropic files for a massive Nasdaq IPO, and Moonshot AI rolls out Kimi K3.1 Pro amid compute constraints.

A new report from METR and Redwood Research reveals critical blind spots and exploit loops in OpenAI's internal agent monitoring systems. We also explore hardware announcements including OpenAI's custom inference chip Jalapeño, Cerebras's multi-wafer deployment architecture, and TypeSafe AI's discriminative model Jev.

Anthropic’s Claude Duet brings persistent, background AI assistance to connected apps—but its approval gates and unseen browser raise questions about how well agents can be monitored. This roundup also covers METR’s evaluation concerns, Recursive’s local coding model Flint, Unitree’s security humanoid, Google and Meta’s TorchTPU, OpenAI’s Orchestra Bio acquisition, and the financing pressures behind the AI infrastructure boom.