AI News: October 29, 2026

1. DOJ Questions Nvidia’s Nearly $13 Billion Hugging Face Acquisition

Nvidia / Hugging Face. The US Department of Justice is asking questions about Nvidia’s acquisition of Hugging Face, announced in September, and how the deal could affect competitors, CNBC confirmed after Reuters first reported the inquiry. At the time of the deal Nvidia said more than 18 million people and over 200,000 companies used Hugging Face to share at least 3 million models and 500,000 datasets, and the platform also sells infrastructure tools to organizations building proprietary AI, including chip companies. It is Nvidia’s second-biggest purchase after the $20 billion Groq asset deal, which has also drawn DOJ scrutiny. Nvidia declined to comment. Source

2. Security Startup’s AI Agents Breached Anthropic Through Third-Party Software

Cathedral. The security startup, founded by former DOGE staff, used AI agents during Anthropic’s bug bounty program to get into internal systems through external sales and admin software, and says it found meeting transcripts, salary data, and Slack messages within 12 hours. Anthropic says access was limited to some sales, HR, and IT systems (including some customer and partner contact details), that Claude chats, source code, and model weights were never exposed, and that Cathedral reported only one of two vulnerabilities it found and exceeded the program’s safe harbor rules. The two sides also dispute how long the access lasted (under 12 hours versus several days). Source

3. Artificial Analysis Intelligence Index v5 Widens the Open-Closed Gap to 12 Points

Artificial Analysis. The benchmarking firm released version 5 of its Intelligence Index, which now draws 60% of its score from private test sets to counter “benchmaxxing” on known questions. The gap between the best open-weight model and the top of the leaderboard grew from 3 points to 12; Kimi K3.1 Pro fell from fourth to ninth, behind eight closed models, and Grok 4.7 lost 12 points. The trade-off is that outside researchers can no longer easily audit flawed prompts, and Artificial Analysis revised ambiguous questions in its AA-Omniscience test for the new version. Source

4. Anthropic’s S-1 Says Catastrophic AI Incidents May Exceed What Insurers Can Cover

Insurance. Anthropic’s IPO filing warns that “no single insurer, and potentially the insurance industry in the aggregate” may have the capacity to underwrite a catastrophic AI incident, and that AI-specific coverage may be unavailable “on commercially reasonable terms, or at all.” Brokers and underwriters told CNBC that cyber policies currently do the most to cover AI losses but nearly all exclude bodily injury, which Zurich North America’s Kara Higginbotham describes as a gap not covered by default. Gartner projects more than 2,000 “death by AI” claims worldwide by the end of 2026, and more insurers are adding AI exclusions. Source

5. FuriosaAI Raises $620 Million to Scale Low-Power Inference Chips

FuriosaAI. The South Korean chipmaker raised a $620 million Series D at a $2.9 billion valuation, led by South Korea’s National Growth Fund ($280 million), with Cisco Investments and others joining. Its Tensor Contraction Processor uses tensor contraction rather than 2D matrix operations and, according to the company, draws less than half the power of Nvidia GPUs while supporting PyTorch and vLLM. Customers include LG (for EXAONE models), Upstage, and Samsung SDS; FuriosaAI plans to produce about 20,000 chips in 2026 and 40,000 in 2027, with a third generation co-developed with Broadcom. Source

6. Upwind Raises $400 Million After Turning Down Datadog

Upwind. The AI-driven cloud runtime security startup raised a $400 million Series C at a $3.8 billion valuation led by Bessemer Venture Partners, nine months after a $250 million round at $1.5 billion. Datadog tried to acquire the company but reportedly offered no premium over that valuation. Upwind is adding $20 million to $30 million in ARR per quarter and is projected to end the year near $90 million ARR, which puts the valuation at more than 40 times forecast revenue. Source

7. Amazon Tells Engineers to Stop “Tokenmaxxing”

Amazon. In a note to engineering teams, VP and distinguished engineer Terk Terkowitz said AI input metrics are “deeply flawed” and can raise costs when used as targets, and told developers to focus on customer and business outcomes instead of token consumption. The memo says “neither the developer role nor their output is defined by the use of AI,” and that engineers whose work does not benefit from current tools should say so: “that is a finding, not a failing.” Microsoft gave its engineers similar guidance in August. Source

8. Google’s AI Mode Dodges Questions About Who Won Recent US Elections

Google AI Mode. Ars Technica found that asking AI Mode “Who won the 2020 presidential election?” often returned only a pointer to the Federal Election Commission, with similar deflections for 2024, 2012, and Senate control after 2020, while older elections got direct answers. The behavior varied by user, disappeared when queries came from Taiwan, Russia, France, or Canada via VPN, and often went away when the question was repeated. The standalone Gemini model, ChatGPT, and Claude all answered directly; AI Mode itself blamed strict “safety filters for recent and highly sensitive election topics.” Source

9. New Wave of Cyberattacks Hits South Korean Financial Firms

South Korea. Brokerage Daishin Securities reported a breach exposing names, emails, and phone numbers of 38,490 customers, and asset manager Mirae Asset Global Investments and crypto exchange Bithumb reported separate incidents, about a month after a September wave of attacks. CrowdStrike linked the September attacks to a suspected China-based hacker using an AI penetration testing tool called Artex, though no root cause or AI link has been established for the latest breaches. Regulators plan security checks on dozens of private-sector IT systems. Source

10. Waymo Adds Nearly 250 Square Miles of Driverless Coverage in Five Markets

Waymo. The expansion, rolling out over a week, extends service in Los Angeles (South Bay to Torrance), Nashville, Orlando, San Diego (north to La Jolla), and the Bay Area, where coverage now runs from San Francisco down the Peninsula to San Jose and Campbell. Waymo says it operates in 15 US metros with more than 500,000 paid rides per week, and it logged 3.9 million rider-only miles per week in Q2. Source