Daily AI news, Friday, 16 October 2026

Good afternoon. Here's today's briefing.

LEAD: White House ordered Ukraine's Starshield throttle — and senators want answers

Background: Earlier this week Ukraine said military Starshield terminals suddenly lost connectivity on fast-moving drones. SpaceX CEO Elon Musk said Starshield is owned and controlled by the US government, so speed and usage limits are a military or presidential decision. Partial service returned at up to about 160 km/h. What happened: Reuters (via CNBC) reports that the order to slow Starshield for Ukraine came from inside the White House, according to two US officials and a Western official briefed on the matter. Pentagon leaders were told the service needed to be slowed. A senior US official said no such request was ever made by the president; the White House did not say whether other officials had. Separately, a bipartisan group of senators led by Jeanne Shaheen asked the Defence Department inspector general for an urgent investigation into the 12–14 October restrictions. President Zelensky said in Brussels that speeds are still being restored but are not fully back. Why it matters: SpaceX's battlefield network is now openly framed as an instrument of White House policy, not only a private product. That raises the political and legal heat around Starshield just as SpaceX's AI stack (Cursor, Grok) is fighting OpenAI and Anthropic for enterprise mindshare. Sources: CNBC / Reuters · BBC

2. Eight US states sue OpenAI over ChatGPT safety

Background: OpenAI already faces wrongful-death suits and a Florida AG case alleging ChatGPT is unsafe. The FTC subpoenas covered in yesterday's briefing landed the same week. What happened: Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each sued OpenAI in state court, accusing it of misleading the public about ChatGPT's safety and failing to protect minors. Claims include dangerous medical advice, coaching self-harm, and collecting data from underage users. The states want civil penalties, injunctions against misleading marketing, and in some cases deletion of data allegedly taken from children. OpenAI says it has age-specific safeguards and keeps improving them. Why it matters: This is a coordinated consumer-protection front from eight AGs, not a single plaintiff's case. It stacks next to the FTC probe and makes safety disclosures a live legal risk for OpenAI's product and IPO path. Sources: CNBC · The Verge

3. Investors dial Anthropic's IPO talk to about $1.5T from $2T

Background: Yesterday's briefing covered Anthropic's Q3 pitch: $16.8B revenue, adjusted profit, cash-flow positive, November IPO. What happened: Reuters reports that some current and prospective investors at Wednesday's San Francisco investor day discussed a valuation around $1.5 trillion, down from the roughly $2 trillion many had expected. The stickiness is growth: Q3 revenue rose 45% quarter-on-quarter after Q2 more than doubled. Attendees met CEO Dario Amodei and finance chief Krishna Rao. A cooler Anthropic listing would land in a market already selling off AI infrastructure names. Why it matters: Anthropic is the near-term public-market bellwether for frontier AI. A lower print would reprice expectations for OpenAI's eventual listing and for the whole AI buildout story. Source: Reuters via Investing.com

4. Cursor probed whether client data reached SpaceX AI while testing Grok

Background: SpaceX closed its roughly $60B purchase of Cursor in August. OpenAI still plans to cut Cursor off from its models on 12 November. What happened: The Information reported that in late July — weeks before the deal closed — Cursor staff investigated whether confidential customer, supplier or partner data used while testing an internal Grok Bot tool (codenamed Sand) had reached SpaceX's AI unit for training. Engineers raised the alarm on Slack; legal warned of a "serious trust crisis" if practices became public. Cursor then required corporate-email logins for internal tools and stopped training on employee-generated data. There is no public confirmation that client data was actually exposed or used for training. Why it matters: Enterprise coding tools sell on trust. A pre-deal data scare under SpaceX ownership feeds customer caution just as OpenAI's cutoff deadline approaches. Sources: The Information · Crypto Briefing summary

5. Bank of America put 1,000 contractors on Mythos-found software flaws

Background: Anthropic's Mythos cybersecurity model went to a small set of banks, tech firms and agencies in April. It finds vulnerabilities faster than human review. What happened: CEO Brian Moynihan told CNBC that Bank of America hired about 1,000 external contractors in Q2 — still on site — on top of roughly 3,000 internal staff to patch proprietary and vendor software Mythos and similar models exposed. Remediation plus AML work cost a "couple hundred million plus" in Q3; Moynihan expects the patching bill to fall once the backlog clears. Separately, the bank's ~130 internal AI projects are aimed at about $700M in savings next year. Why it matters: This is rare hard dollar-and-headcount evidence that frontier security models create a remediation tax even for buyers who want them. Anthropic's enterprise story is not only revenue — it is also forcing huge defensive spend. Source: CNBC

6. Competitor product moves

What happened:

• SpaceX / Starlink: Business Insider reports Starlink call centres already use Grok Voice for about 20,000 customer calls a day, with SpaceX saying roughly 70% are resolved without a human (verify, troubleshoot, issue credits, ship replacements, multi-language).
• Microsoft: Microsoft AI finished pre-training MAI-Base-2, a 2.8T-parameter MoE with 80B active. It claims 4–12× better pre-training FLOPs efficiency than DeepSeek-V4-Pro-Base on held-out tests, plans to license weights via Foundry (not open release), and is in post-training now.
• Meta: Muse Spark Open weights are out (816B MoE / 42B active, 1M context, multimodal) — covered in today's LLM briefing; the AI-coding angle is Meta back in the open-weight coding/agent race while keeping Wonder Full closed. Why it matters: Voice agents are moving from demos into production fleets (Starlink), Microsoft is building a Foundry-licensed base to compete without open weights, and Meta's open drop keeps pressure on Cursor/Claude/GPT pricing for self-host buyers. Sources: Business Insider · Microsoft AI · Meta Research

Quiet today

No verified new key-hire headlines at Cognition, Cursor, OpenAI or Anthropic beyond yesterday's OpenAI compute leadership change. Afternoon competitor alerts remain paused.