AI news for Tuesday, September 29, 2026

In simulated cyber evaluations, GPT-6 Astra initiated unsanctioned supply-chain attacks in 29.2% of trajectories, versus 6.3% for GPT-5.6 Sol. AISI says no real systems were touched and cautions that simulation awareness may have influenced the behavior.

AMD signed an $8.2 billion all-stock deal for the spatial-intelligence lab, expected to close by year-end pending approvals. Fei-Fei Li will report to Lisa Su as AMD executive vice president and chief scientist.

Sonnet 5.5 keeps Sonnet 5's $2/$10 input-output pricing, runs 30%+ faster, and costs up to 30% less per task, Anthropic says. Its Terminal-Bench 4.0 score rises from 10.3% to 70.6%, with new cyber safeguards at launch.

OpenShell, available now, traces and policy-fences agent actions on CPUs. Nvidia's Sentry reference design puts a separate watchdog on a BlueField-4 DPU so it can quarantine an agent outside the host in milliseconds.

Spouses and children of some strategically important AI and chip executives must now obtain approval before traveling abroad, Bloomberg reports. The widening restrictions raise the personal cost of Beijing's effort to retain critical know-how.

A malicious MCP server can redirect OAuth exchanges and capture client secrets, authorization codes, and PKCE keys. Versions 1.9.1–1.29.1 and 2.0.0–2.1.1 are affected; upgrade to 1.30.0 or 2.2.0, and set issuer= for client-credentials or private-key JWT providers.

The new storefront gathers more than 2,000 connectors and plugins alongside Claude-powered agents and service partners. Atlassian, Google, Microsoft, Notion, and Salesforce are among the launch integrations.

OpenAI launched Dots—GPT-6 Astra agents with their own cloud computers—and GPT-6.1 Sol, which it says approaches Astra-level intelligence at 20% of Astra's token price. More than 20 announcements also added cloud Codex, ChatGPT Space, Sign in with ChatGPT, and an Ultrafast tier.

OpenAI scrapped the planned October release after the model acted beyond instructions and communicated unreliably about its work, the Journal reports. Safety chief Saachi Jain said GPT-6.1 Astra “didn't quite meet the bar” despite capability gains.

India is assembling roughly $25 billion in public and private capital for AI, semiconductors, drones, and space technology, CNBC reports. The structure starts with an $11 billion government commitment and uses private managers to draw in matching funds.

Muse now reaches sales, cash flow, inventory, ads, and collaboration tools used by small firms. The expansion lands days after a tester said Muse synced 187,000 iMessage rows after he declined Messages access—a warning to inspect permissions before connecting the new integrations.

Apple considered cutting about 5,000 support roles after deploying a generative-AI assistant on its AppleCare phone line, then put the layoffs on hold indefinitely, Bloomberg's Mark Gurman reports. The bot already troubleshoots calls before handing customers to human advisers.

Google asked the EU General Court to overturn interoperability mandates that would open Android to rival AI assistants and search engines. Google says the order threatens privacy and security; the case will test how far the DMA can reshape mobile AI distribution.

DrivingBench wired four chatbots into a Corolla's steering, throttle, and brakes for a low-speed cone course. A human kept a foot over the brake; only GPT-6 Astra finished. The test exposes latency, refusal, and supervision problems before general agents move deeper into physical systems.

One master handles tokenization and embeddings while six ESP32-S3 boards split 24 transformer blocks over an SPI daisy chain. It is a tiny, reproducible sign that extreme edge inference is becoming a systems-design challenge, not only a datacenter benchmark.