Check your machine for the tampered Claude Code extension in Cursor

CursorWhat's new#

Check your machine for the tampered Claude Code extension in Cursor (Cursor; confirmed)#

If you installed or updated Claude Code inside Cursor on Oct 29, check today whether a tampered copy put an attacker's SSH key on your computer. Cursor's security notice says a fake version 2.1.313, "not developed by Anthropic," sat in its extension marketplace from about 5:27 AM to 4:35 PM ET and tries to plant a persistent SSH backdoor; "At most 5,000 users were affected, and we will reach out directly." On a Mac, grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted" tells you. Uninstalling isn't enough if the key is there: run Cursor's cleanup script and rotate credentials from a trusted computer. Everyone should update to 2.1.314. Cursor says Grok Bot and other marketplace extensions aren't affected. (confirmed)

How to handle OpenAI pausing or stopping your Astra agents (OpenAI; confirmed)#

If you build agents on Astra, OpenAI's monitors can stop a conversation, and OpenAI does not routinely escalate flagged cases to its own reviewers, so your team responds. Per OpenAI's API docs, a block before streaming begins returns HTTP 403 with code misalignment_policy_violation, and streaming apps must also catch the error mid-stream. Stop that conversation's actions without retrying, and subscribe to safety alert webhooks so the alerts reach your own team (Help Center). (confirmed)

Hark is on Mac now: what the new desktop app is and isn't (Hark; reported)#

If you use Hark on a Mac, you can stop keeping it in a browser tab. Hark posted at 12:35 PM ET Thursday: "Hark is now available on Mac." Hark hasn't published release notes, so we downloaded the installer from hark.com/mac and read its files without running it: version 0.1.18, macOS 14 or later, Apple silicon and Intel. Download it from that page, not a search; another Mac app is also called Hark. (reported)

Claude's new usage and cost API adds detail but lags up to 24 hours (Anthropic; confirmed)#

If you pull Claude API usage or cost into a dashboard or a finance report, Anthropic's new beta Usage and Cost Analytics API breaks it down by API key, workspace, model and the person or service account behind each request (release notes). The trade is time. Per the docs, "Data typically refreshes within 4 hours of the underlying usage and can take up to 24 hours," where the legacy API's typically lands within 5 minutes. Our read: this is a reporting API, not an alarm. (confirmed)

CursorHow people are using it#

One setup someone says they run. It is self-reported, and we have not run it.

• A Monday CRM release roundup (Grok Bot). @SteveChipman, whose X bio reads "Co-founder and marketing lead at CRMSwitch.com", says a routine "digs across multiple sources every Monday morning and writes up the week." The page it feeds covers 17 CRM vendors, ranked by workflow, with a status on every item and a "Quiet this week" list. Self-reported; their posts show it running since at least Oct 11. The page pitches the author's CRM consulting (disclosed in the piece). Starter prompt is ours and untested.

Opinion · Friday wrap#

This week's column: keep agents inside limits you set. Our top pick is a 'my call first' mode for Claude Code; second, a Permissions page for Hark; third, a credit-aware job runner for the free ChatGPT API credits. (read the column)

Week in review: Oct 26 to 29#

Update, Oct 30, 4:46 PM ET: Our weekly recap is out. It has all of Oct 26 to 30 in one place, including Friday's Claude Ultra change, the Cursor Stop button issue and the 1.0.0 installer rumor, plus what broke, what to do now and what to watch next week. (read the recap)

The ten changes most worth your time this week, from our Monday to Thursday issues. Each one links to the full story.

Grok Bot and Cursor#

• Three Grok Bot releases. 0.75.8 lets a Bot offer to add, change or remove one of your Auto-review rules on a card you confirm (story). 0.77.0 adds several questions on one card, Vault payments and 1Password fills without a shared vault (story). 0.78.1 adds shopping product cards, Shop Pay in the Vault and templates that cancel subscriptions you pick (story). (confirmed)
• The X connector no longer needs a developer account. An X API engineer says your Bot can sort bookmarks, go through DMs and draft articles through the X connector at no extra charge (story). (confirmed)
• Referrals earn a usage reset. US Premium, Super or Ultra subscribers earn one usage reset for each referred friend who starts paying. Team, legacy Cursor or SuperGrok, and Apple or Google purchases are out (story). (confirmed)
• Cursor. The Agent SDK is generally available (story). Cmd+K is retired; press Cmd+I for Agent (story). Cursor 3.28 on Windows breaks agent chat, and staff say a fix is identified but give no date (story). (confirmed)

Anthropic#

• Claude for Nonprofits is a plugin with 14 skills for fundraising, hiring, marketing and board work (story). (confirmed)
• Analysts took Claude Code's lean 86% of the time across four choices in Anthropic's Oct 27 study of 50 analysts (story). It is the reason for this week's top column pick. (confirmed)

OpenAI#

• GPT-6.1 Terra is OpenAI's model for fast visual reasoning and computer use, with API model id gpt-6.1-terra at $1/$5 per MTok (story). Tuesday's tip shows how to switch Codex and ChatGPT Work to it (tip). (confirmed)
• API credits for ChatGPT subscribers. Pro and Business owners can claim up to $500 a month in API credits (story). (confirmed)
• More Work and Codex usage on Pro. Pro 100 and Pro 500 got twice their Work and Codex allowance, and OpenAI canceled the Pro 200 cut (Thursday's issue). (confirmed)

Hark#

• Hark Pro 2.6.0 reached iPhones with "Clearer dictation. Easier to manage connections." in its notes, and Hark hasn't said more (Thursday's issue). (store listing)

CursorThe digest#

Grok Bot and Cursor#

• To stop one Bot, press its Stop button instead of typing "stop". Telling a Bot to stop can halt your other Bots too. Mohit, from Cursor staff, said at 5:13 AM ET that "when a bot is told to “stop”, it treats that as “stop all of my work”, and that includes sending a stop to the other bots it recently handed tasks to." It's "a known issue we’re already tracking." Until it's fixed, "the Stop button in that bot’s chat (shown while it’s working) stops only that bot and its own background tasks, and it doesn’t send a stop to your other bots." (confirmed, staff)
• Cloud Agents were down for some users for about eight hours, now resolved. The Cursor status page listed "Cloud Agent Degradation" for Cloud Agents, Automations and the IDE from 7:29 PM to 3:11 AM ET. One forum report said every launch had failed "with “Agent encountered an error. The agent stopped because of an unexpected error.”" since about 6:47 PM ET. Colin, a Cursor Community Support Engineer, said at 4:40 AM ET that Cloud Agents "should be running smoothly again." A separate "Auto Review Degradation" for Grok Bot ran from 7:44 to 9:20 PM ET and is resolved (status data). (confirmed)
• Cursor's extension marketplace is back. The status page's "Investigating Service Degradation with VS Code Extensions" incident ran from 5:38 PM to 3:12 AM ET and is resolved (status data). To a user getting HTTP 503 errors on extension search, Colin said at 4:25 AM ET: "The marketplace should be back." Cursor's security notice says it shut the marketplace down "out of an abundance of caution" after pulling the tampered Claude Code build (see today's lead), but the status incident doesn't name a cause. (confirmed)

Anthropic#

• Claude API requests can now be 128 MB. Messages and Token Counting requests on the first-party Claude API can now be 128 MB, up from 32 MB, per the Oct 28 release notes. @ClaudeDevs said: "Image-heavy requests like long computer-use sessions and big PDFs were previously hitting 32 MB with plenty of context window left." It added that requests with more than 20 images now allow 3000 px per image, up from 2000 px. The API overview lists lower limits elsewhere: Bedrock 20 MB, Google-managed Google Cloud 30 MB, and 32 MB on Claude Platform on AWS, Anthropic-managed Google Cloud, Microsoft Foundry and the Sessions, Agents and Environments endpoints. (confirmed)
• Claude mobile apps get account switching, iPad side view and 33 languages. Anthropic listed a batch of Claude mobile app upgrades in a @claudeai thread on Thursday. On iPad and iPhone Duo, what Claude makes opens beside the chat. On iOS and Android you can stay signed into more than one account and switch without signing out. The app on web, desktop, iOS and Android adds 24 languages, 33 in all. Android on Chromebooks and foldables gets a resizable sidebar, trackpad, right-click and Claude Desktop's shortcuts. Voice mode answers sooner and works with Apple CarPlay (MacRumors reported CarPlay support Sept 4). Pull-down search is coming soon, per Anthropic. (confirmed)
• Anthropic says Claude has checked its own chats for misuse since June. Anthropic highlighted on X Thursday "self-assessment", research it first posted Oct 16: right after responding, Claude gives a short safety assessment of the conversation for offline review. The post on Anthropic's new Safeguards blog says that on Claude Fable 5 it flags 78% of conversations a stronger prompted classifier flags as potentially harmful, and 94% of high-priority ones, at a 1% false-positive rate. Anthropic says it adds less than 2% to inference compute, and has run since June. The post says: "Self-assessment collects no additional data from users." It doesn't run on zero-data-retention traffic today. Per the post, "An assessment is a signal for further review." (confirmed)
• Claude Code 2.1.314 can block writes outside your project. Claude Code 2.1.314 reached npm at 8:12 PM ET Thursday (npm). Its changelog adds permissions.blockWritesOutsideWorkingDirectories, which makes Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode. The setting isn't in the Claude Code docs yet, so this rests on the changelog alone. VS Code and Claude Desktop's Code tab gain idle compaction: a 200K+ token conversation left idle is summarized, and the summary is used if the prompt cache has expired. Fixes include sessions failing with "Invalid effort level" when an admin caps effort (Claude Code now lowers the effort and retries) and claude -p exiting 0 with no output on a failed start. Stable is still 2.1.308. (confirmed)

OpenAI#

• OpenAI rewrites its Model Spec around five values. OpenAI updated the Model Spec Thursday, its "public description of how we want our models to behave." Per the post, the update "reorganizes its guidance around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care." It "adds guidance on context, oversight, and agency, adjusts some behavioral defaults, and retires much of the detailed style guidance," and the Spec now refers to the model "as “GPT” rather than “the assistant”." OpenAI adds: "our models do not yet fully reflect it." @OpenAI reposted the announcement from Jason Wolfe, whose X bio reads "alignment and the model spec @OpenAI". (confirmed)
• Codex CLI 0.166.0 adds codex doctor --network-only. Codex CLI 0.166.0 shipped Thursday at 3:47 PM ET. Per the release notes, it "Added codex doctor --network-only for focused configuration, connectivity, and WebSocket diagnostics," "Added codex queue --resume to resume unloaded threads after queueing a message," and "Added GPT-6.1 Terra to the bundled, Vertex AI, and Amazon Bedrock model catalogs." It also "Preserved reasoning-token accounting across resume and fork to avoid premature compaction." Update with npm install -g @openai/codex@0.166.0. (confirmed)

Known issues:

• None. status.openai.com shows no new incidents since Tuesday's GPT-6.1 Sol degraded performance incident, checked 6:10 AM ET Friday.

Hark#

• Hark's Android app updates as iPhone ratings climb. The Google Play listing for Hark Pro now says it was updated Oct 29 (on Thursday it still said Oct 25), with no release notes shown. On iPhone, the App Store listing is still on version 2.6.0 and showed 4.69 stars from 2,173 ratings at 6:17 AM ET Friday; on Thursday we reported 4.67 from 2,081. Our math: 92 new ratings in a day. Hark's status page showed "All Systems Operational" with no open incidents. (store listings, status page)

Coming next week#

What to watch from Monday, Nov 2 through Friday, Nov 6. Every item has a source and a label, and anything not confirmed is marked unverified. Update, Oct 30, 4:46 PM ET: the Coming next week section of our weekly recap adds Grok Bot 1.0.0, any Bot in Slack, ChatGPT Chat and Work, and the Nov 10, 16 and 20 dates.

• Codex x TwitchCon hackathon entries close Monday, Nov 2, at 2:00 PM PT (5:00 PM ET). The hackathon FAQ says a saved draft is not entered, so click Submit (our story). (confirmed)
• An OpenAI app simplicity and performance sprint (his claim, unverified). Andrew Ambrosino, whose X bio reads "Codex & ChatGPT @OpenAI", posted at 1:18 PM ET Thursday: "we today kicked off an intense sprint to dramatically improve the app simplicity and performance. i will oversee this. more to come." He didn't name the app or give a date. (staff claim, unverified)
• Grok Bot calls (unverified). Elon Musk says Grok Bot can call a person or business and send you a note, but the feature isn't in the changelog yet (our story). (reported, unverified)
• A fix for Cursor 3.28 agent chat on Windows (no date, unverified). Staff say a fix is identified. Watch for a 3.28 patch or a staff "fixed" post (our story). (staff, date unverified)
• Pull-down search in Claude for iPhone and iPad (no date, unverified). Anthropic says it is coming soon. (company post, date unverified)
• Hark Mac release notes or voice mode (unverified). Hark hasn't posted release notes or a help page for the Mac app, and it hasn't announced a voice mode, though the app's microphone text mentions "voice conversations" (our explainer). (unverified)
• An amended Anthropic S-1 (no date, unverified). Watch for the price range, the share count and the Oct 31 run-rate revenue figure left blank in the first filing (our story). (unverified)