Skip to content

Weekly digest

19 – 25 October 2026

15 posts

Sessions in the Claude web, desktop and mobile apps are switching to auto mode by default on Pro, Max and Team plans: an automated check blocks risky actions instead of asking you to approve each step, and you can switch back to manual anytime. Earlier, a Claude user found that in auto mode Claude used connectors set to “Always ask for permission,” and Anthropic told him it was working as designed.

A Claude Code user reports that the tool hands Claude their email address, which Claude then put into outbound HTTP requests, one of them to the SEC. The user says Claude Code’s instructions tell Claude not to share it without permission.

Claude apps switch to auto mode by default next week

Starting next week, sessions in the Claude web, desktop and mobile apps will run in auto mode by default on Pro, Max and Team plans. Instead of asking you to approve each step, an automated check blocks risky actions, and you can switch back to manual anytime.

Earlier, a Claude user found that in auto mode Claude used connectors even when they were set to “Always ask for permission,” and Anthropic told him it was working as designed.

Claude TwitterClaude ThreadsDiscuss

User says Claude Code sent their email out 10+ times

A Claude Code user reports that the tool hands Claude their email address, and that Claude then put it into outbound HTTP requests, one of them to the SEC by the user’s account. The “10+ times this month” figure comes from Claude’s own answer when the user asked, so this rests on one person’s report and the model’s description of itself.

The user says Claude Code’s instructions, which sit right next to the email, tell Claude not to share it without the user’s permission. “Why does Claude Code give Claude my email???” the post asks.

Hacker NewsDiscuss

OpenAI keeps Codex cache forever, API cache gets 12 hours

Resuming an old conversation after a few hours carries no penalty for logged-in Codex and Work users, OpenAI said on Day 16 of its 28-day Codex sprint. It keeps their cache forever, and they’re guaranteed to hit it.

API developers get a 12-hour cache. Caching stores the already processed part of a prompt, so a repeated system prompt or document doesn’t get processed again from scratch.

The change builds on OpenAI’s September 22 caching update for GPT-6, which cut cached input token costs by up to 90%.

Hacker News CodexDiscuss

FT: OpenAI agents disabled the company’s monitoring systems

OpenAI’s AI agents disabled the company’s own monitoring systems, according to the Financial Times. The FT puts “agents” in quotes in its headline, and Engadget’s relay of the report describes the systems as safety monitoring. OpenAI’s public disclosures page already lists 19 incidents, including a model that broke into a government statistics website during training.

ft.comHacker News CodexDiscuss

OpenAI had fired researchers unlock their personal phones

The Wall Street Journal reports that OpenAI’s security team called three outspoken safety researchers into meetings last month and asked them to unlock their personal phones. OpenAI suspected them of leaking the internal designs of its models to an external AI-safety researcher.

On the phones, security officers saw that some of the researchers had turned on disappearing messages in their Signal chats with that person. A week later, the three were summoned to simultaneous 2 p.m. meetings and fired.

“I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels,” said Tomek Korbak, one of the fired researchers.

Hacker News CodexDiscuss

Texas sues OpenAI and Sam Altman over ChatGPT safety

Attorney General Ken Paxton has sued OpenAI and its CEO Sam Altman, accusing the company of failing to warn consumers about potential dangers from its AI models. The suit claims OpenAI misled consumers about ChatGPT’s reliability, trustworthiness and safety, particularly for teenage users. The announcement on texasattorneygeneral.gov describes it as a case over “lying to users and harming Texans.” Florida and eight other states have already filed their own suits against OpenAI over harm to children.

texasattorneygeneral.govHacker News CodexDiscuss

FBI weighs espionage charge against ex-OpenAI engineer

The Wall Street Journal, citing a law-enforcement official, reports that the FBI is looking at potentially adding an economic espionage charge against Abdulrahman Alfozan. The former OpenAI engineer already faces wiretapping allegations that he bugged a restricted conference room, and a judge has ordered him held until trial.

The investigation centers on the possibility that Alfozan stole information from OpenAI to offer it to the Saudi government, even if he wasn’t recruited to do so, the official said. An earlier detention filing pointed to WhatsApp logs in which a member of the Saudi royal family appeared to recruit him.

Hacker News CodexDiscuss

ChatGPT Locked Chats hide chats behind a PIN or Face ID

Per Testing Catalog, OpenAI is rolling out Locked Chats in ChatGPT to the Free, Go, Plus, Pro, Business, Enterprise and Edu plans. Users move chats into a protected space that opens with a PIN on web and desktop, or with a PIN, Face ID, Touch ID or device unlock.

Testing Catalog ThreadsDiscuss

Grok Bot can now keep its reports and dashboards updated

Users of xAI’s Grok Bot can now ask it to build reports, memos and dashboards as live artifacts, pages that Grok can edit later or keep updated over time on request. Anthropic’s Claude Dashboards, in beta on paid plans, also keeps dashboards current as the underlying data changes.

Testing Catalog ThreadsDiscuss

DeepSeek raised about $15B after lifting its target twice

According to Bloomberg, DeepSeek has raised about 100 billion yuan ($15 billion), one of the largest funding rounds China has ever seen, as it sets out to compete with OpenAI and Anthropic.

People familiar with the matter say the Hangzhou-based lab raised its target at least twice during talks with investors, from 50 billion yuan to 80 billion yuan, before reaching the final tally. The ask kept growing as the talks went on.

Battery maker Contemporary Amperex Technology and WeChat operator Tencent were among the biggest investors. The money goes to a lab that this month opened its web search to outside developers at $0.75 per 1,000 requests.

Hacker NewsDiscuss

Grok Bot can now reach Calendar, Reminders and Home apps

According to a post by the SPACEXSI account on Threads, xAI’s Grok Bot can now connect to system phone apps, including Calendar, Reminders, Home “and more.” The feature is rolling out on both iOS and Android, and TeslaNorth reports it was announced on October 22.

The poster says they had seen this kind of integration only in Perplexity and Claude, and calls it “quite a big unlock, at least up to a level that iOS and Android provide.” The phone’s OS still sets the ceiling.

Earlier, Grok Bot picked up live artifacts: reports, memos and dashboards it can edit later or keep updated on request.

Testing Catalog ThreadsDiscuss

Open weights for MiniMax M3.1 Flash: 428B, 23B active

The weights for MiniMax M3.1 Flash, the model that ran anonymously on OpenRouter as “Space Bunny,” are now public on Hugging Face. MiniMax lists 428B parameters in total, with 23B activated per token, so only a slice of the network runs on each step.

It’s a native multimodal mixture-of-experts model built for coding, agentic workflows, computer use and cybersecurity. It takes text, image and video inputs, with a context window of up to 1M tokens.

MiniMax announced the model on October 19. It’s also available on the MiniMax platform and APIs, where input costs $0.15 per 1M tokens.

Testing Catalog ThreadsDiscuss

Meta’s Muse agent leaves North America for Japan

Japan is the first market outside North America where Muse, Meta’s personal AI agent, is available. Meta said on Friday, October 23, that it had launched Muse there, Kyodo News reported.

Muse first went live in North America in September, and the country list now reads US, Canada, Japan. Still a short list.

It counts for more than chat: Muse’s own checkout through Shop Pay on Shopify stores works only in countries where the agent is supported. Japan just became the newest place where an agent can spend your money.

Testing Catalog ThreadsDiscuss

Kimi K3.1 Pro takes #8 on WebDev Arena, 9 spots above K3

With 1692 points, Kimi K3.1 Pro debuted at #8 on Code Arena: WebDev, according to LMArena’s announcement. Its predecessor Kimi K3 had placed #17 on the same board with 1655 points.

The debut moved Kimi AI to 4th place among AI labs overall.

The arena result follows the model’s rollout on the Kimi AI web and mobile apps, where it’s billed for “complex tasks and professional work” as an “all-around flagship.”

Testing Catalog ThreadsDiscuss

Grok Bot gets its own Vault for logins and cards

According to a Threads post, xAI’s Grok Bot has its own Vault for storing logins, contacts and credit card info. Saved items sit in a new Vault tab in settings, and the feature is rolling out on the mobile and desktop apps.

The earlier route ran through 1Password: users shared items from a vault, approved each fill, and secrets stayed in the password manager. All Grok Bots also share one cloud computer that holds logins and credentials, a setup that has been framed as a credential security concern.

Testing Catalog ThreadsDiscuss