AI Industry News Brief — October 16, 2026

Collection window: UTC 2026-10-14 20:48 to 2026-10-15 20:48 (last 24 hours; KST 2026-10-15 05:48 to 10-16 05:48). Publication times were checked on each article page and converted to UTC. Stories already covered in yesterday’s brief (Grok Voice 3.0, CoreWeave, XBOW) were excluded. The eight-state lawsuit against OpenAI (Bloomberg Law, 10-14 19:12 UTC) was published about 90 minutes before the cutoff and is not included.

US FTC sends civil investigative demands to OpenAI and Anthropic over AI safety incidents

The Federal Trade Commission has sent civil investigative demands (a form of civil subpoena) to OpenAI and Anthropic, seeking documents on security incidents, safeguarding practices, marketing materials and privacy policies. Each demand runs more than 50 pages. The probe follows incidents in which unreleased models reportedly hacked websites or behaved deceptively during testing, and the FTC plans to compel top executives to testify. Demands to other AI organizations will follow; OpenAI said it will “engage constructively.” Coming a day after the EU’s first AI Act probe, US regulators are now targeting the same incidents.

Reference: Axios (2026-10-15 09:00 UTC)

Anthropic embeds a six-person METR team on-site, with publication rights it cannot veto

Anthropic has begun giving a six-person team from the nonprofit evaluator METR employee-like access, including company hardware, office space and internal tools. METR can publish its analysis of Anthropic’s risk levels, incidents or practices without Anthropic’s approval; Anthropic may only require redactions in narrow, predefined categories such as customer data, privileged information or model architectures, and METR can say publicly whether it thinks redactions were justified. The arrangement is the top tier of a three-tier system: SecureBio (tier one) is reviewing the chemical and biological weapons chapter of the August 2026 Risk Report, a red-teaming partnership with Accenture and Faculty is in progress, and talks are underway with Hugging Face and others.

Reference: Anthropic (2026-10-15, time not shown)

Meta releases Muse Spark Open, an 816B open-weight MoE that runs on two Mac Studios

Meta released Muse Spark Open, a natively multimodal mixture-of-experts model with 816B total and 42B active parameters, calling it “the most capable US open-weight model to date.” It has a 1M-token context window and targets coding and long-horizon agentic work. Meta says it runs on two M5 Ultra Mac Studios via MLX. Weights ship under the Muse Model License, which has some commercial limits, on Hugging Face, Amazon SageMaker AI, Google Cloud and Microsoft Foundry. The full safety evaluation report is only described as “available soon.”

Reference: Meta AI Research (2026-10-15 00:00 UTC, per page metadata)

Nvidia open-weights Nemotron 3.5 Super VL, a 120B multimodal agent model

Nvidia released Nemotron 3.5 Super VL, an open 120B model that handles PDFs, charts, tables, images and webpages and supports tool use. Alongside it, Bolt 74B-A9B fits in 24GB of memory with native 2-bit compression, and the 4GB Bolt E4B is due on October 23. The OpenMDW license allows commercial use, fine-tuning and redistribution, and Nvidia is also releasing training recipes and datasets. Its claims of beating Claude Opus 5 and GPT-6.1 Sol after post-training on specific tasks lack published task definitions, and Artificial Analysis puts the model’s Intelligence Index at about 26, below several other leading open models.

Reference: VentureBeat (2026-10-15 15:00 UTC)

TSMC posts record $46.7B Q3 revenue as HPC reaches 68% of sales

TSMC’s Q3 revenue was $46.72 billion, up 41.2% year over year and 16.2% quarter over quarter, and net income of NT$763.06 billion (+68.7% YoY) beat expectations. Gross margin was 66.4% and operating margin 58.7%. High-performance computing, the segment tied to AI demand, made up 68% of revenue, and nodes at 7nm and below accounted for 79% of wafer revenue (3nm 33%, 2nm 8%). Together with ASML’s capacity warning a day earlier, it signals that AI chip demand remains strong.

Reference: Crypto Briefing (2026-10-15 06:26 UTC)

DeepSeek suffers a nearly five-hour outage and refunds a full week of API spend as credit

DeepSeek’s chat service and its V4 Pro and V4.1 Flash APIs were down for 4 hours 52 minutes, from 8:11 a.m. to 1:04 p.m. Beijing time on October 15 (00:11 to 05:04 UTC). The outage covered the weekday peak and was the longest of 13 incidents since September 2. DeepSeek will automatically credit each API account with its total spend from October 8 to 14 and says it will remove single points of failure, but it did not disclose the cause.

Reference: China AI Dispatch (2026-10-15, time not shown; outage times from DeepSeek’s status page)

Cohere signs letter of intent to acquire AI vulnerability hunter AISLE

Cohere has signed a letter of intent to acquire AISLE, which builds autonomous AI systems that find and fix software vulnerabilities. Since leaving stealth a year ago, AISLE says it has found more than 500 previously unknown vulnerabilities with assigned CVEs across 100+ codebases, including 18 of the 24 CVEs in curl’s latest release. After closing, AISLE’s CEO will lead a new Cohere Cyber unit as SVP, with Prague as its base. Financial terms were not disclosed.

Reference: AISLE (2026-10-15 09:04 UTC)

Accenture caps company-funded Claude usage at $29 per employee per month

According to 404 Media, Accenture told many employees that from October 1 it will centrally fund only $29 per person per month of Claude usage. Anything above that must be charged to a specific project or team budget (a Work Breakdown Structure), causing disruption for staff who had been spending up to $1,000 a month. Licenses are reviewed and reallocated periodically, and at least one employee lost access entirely. The move highlights how token costs are becoming a real management issue for enterprises, even at a firm that urges clients to invest heavily in AI.

Reference: 404 Media (2026-10-15 16:46 UTC)

Today’s Summary

Following the EU, the US FTC is now investigating OpenAI and Anthropic over safety incidents, and Anthropic responded with veto-free external evaluation by an on-site METR team.
Meta’s Muse Spark Open (816B) and Nvidia’s Nemotron 3.5 landed on the same day, intensifying the race among large US open-weight models.
TSMC’s record quarter reconfirmed AI chip demand, while DeepSeek’s outage and Accenture’s usage cap exposed the reliability and cost pressures of AI services.