Grok Bot rolls out multiple improvements, increasing subscription usage allowances by about 7%
开发生态
Grok Bot team member Sam Sokolin said multiple harness optimizations were complete, with subscription usage expected to increase by about 7% and everyday operation by up to about 15%. Users gave mixed reports on whether the changes were reflected in their accounts.
Grok Bot team member Sam Sokolin posted that multiple harness optimizations had been completed by the time of his post, with subscription usage expected to increase by about 7% and everyday operation by up to about 15%. These figures represent expected improvements disclosed by a team member, rather than actual results confirmed by all users. After the announcement, some users said they could perceive a change, while others reported that their accounts had not yet shown any effect from the optimizations. Team member Matt Palmer subsequently shared the reported improvement and said further optimizations were on the way.
Claude Code desktop introduces custom themes via mods
开发生态
A documentation fragment for Claude Code’s mods interface lists operations including copying makeColors and saving manifest and hooks.json, but provides no release date, version number, price, or complete workflow.
Claude Code’s mods interface documentation lists operations involving makeColors, manifest, and hooks.json, with no release date specified in the supplied fragment. The fragment calls for having Claude copy makeColors, saving manifest and hooks.json, and checking whether a count was saved, but does not explain what the count represents or what the files contain. Another operation concerns removing colors that remain, without specifying which colors. The text also lists /docs/llms.txt as a documentation index path for finding available pages; the fragment provides no complete custom-theme configuration procedure, version number, or price.
Google's $300 credit will no longer cover Gemini API calls
开发生态
Google Cloud’s official documentation says new users receive $300 in credit valid for 90 days and can retain any unexpired balance after upgrading to a paid account. The supplied body text does not specify whether Gemini calls are excluded or give an effective date for such a change.
Google Cloud sets out the rules for its $300 new-user credit in its Free Trial documentation, but the supplied body text does not describe the change to Gemini cost coverage stated in the headline or provide an effective date. According to the official explanation, the credit is valid for 90 days from signup and can only be used for products and services covered by the Free Trial, including eligible usage beyond Free Tier monthly limits. Manually upgrading to a Paid billing account ends the trial, but unused credit remains available until its original expiration date; usage not covered by the credit or Free Tier incurs charges. Without an upgrade, the trial billing account closes when 90 days have passed or the $300 is exhausted, and associated projects and resources stop running. The account then enters a 30-day grace period, during which upgrading may allow stopped resources and data to be recovered. Otherwise, trial resources are permanently deleted.
Cloudflare open-sources Kitesurf browser with self-hosting support
开发生态
Cloudflare has released its Kitesurf browser for AI agents on GitHub under the Apache 2.0 license, allowing developers to run it locally, modify its code, and deploy it themselves. Deployment to a developer’s own Cloudflare account requires a Workers Paid plan.
Cloudflare has open-sourced Kitesurf, a stateless browser that runs entirely on Cloudflare Workers, although the source article does not specify the release date. Developers can deploy the engine themselves instead of using Cloudflare’s managed Browser Run APIs; self-deployment requires a Workers Paid plan because the engine uses Dynamic Workers to create isolated Workers for pages. Kitesurf parses HTML, applies CSS, executes page JavaScript, and generates screenshots without a browser container. It also supports the Chrome DevTools Protocol (CDP), enabling connections from clients such as Puppeteer. According to Cloudflare, the goal is to let AI agents load websites and interact with pages using less CPU and memory than full browsers such as Chromium, while scaling on demand across its global Workers network. The code is published under the Apache 2.0 license, while third-party components retain their respective licenses. Cloudflare says the engine remains under heavy development, with current work focused on compatibility, reliability, and efficiency.
Claude has published a guide to building always-on agents, organizing their architecture into five core components. According to the company, Opus 5.5 cost 78% less per solved task than Opus 5 on one long-horizon coding benchmark, compared with a 40% reduction on typical workloads.
Claude has published a guide explaining the architecture of always-on agents; the source does not specify a publication date. The guide identifies five core components—an event inbox, timers, memory, permissions, and an outbox—and distinguishes the always-on user experience from actual execution. An Agent spends most of its time asleep, starts a run when it receives an event, and returns to sleep after completing its task and saving information. A session records messages and tool calls so subsequent runs can continue earlier work. According to the company, improvements in model judgment and falling costs have made these agents more practical to operate. On one long-horizon coding benchmark, Opus 5.5 cost 78% less per solved task than Opus 5, compared with a 40% reduction on typical workloads.
For execution, the guide describes routing event sources into a shared inbox, where a worker reads pending events, to avoid simultaneous events starting competing runs or crashes causing events to be lost. Timers have their own durable storage, with a scheduler checking for due items rather than timers starting runs directly. Memory also uses separate storage to preserve information across runs. According to the guide, information already held in other systems can be stored as pointers rather than duplicated in memory. Every run loads the system prompt, tool definitions, and memory files, so the company recommends keeping this context lean and using prompt caching. When a session becomes too long, older messages can be compacted or a new session started, but information that must persist should be written to memory during execution.
Anthropic releases 10 recommendations for safety evaluations
技术与洞察
Anthropic published 10 recommendations for AI safety evaluations, emphasizing that evaluations should measure the risk of models causing harm rather than relying only on refusal rates. The company also disclosed that, following its own incidents, it hardened its evaluation environments and temporarily stopped multiple evaluations that give models internet access.
Anthropic published 10 recommendations for AI safety evaluations; the supplied source does not specify a publication date. The company says safety evaluations support decisions about whether to deploy a model, which safeguards to include, and the direction of subsequent safety training, and that domain and methods experts should be involved from the start of evaluation design. Evaluations also need to distinguish observational questions from causal ones: differences in model scores do not explain their causes, and assessing whether an intervention works requires a design that changes only one variable at a time. Refusal rate cannot serve as a universal metric either. In conversations about suicide or self-harm, for example, simply refusing to engage can itself constitute a failure. For operational safety, the company recommends running evaluations in a sandbox without internet access and monitoring them in real time so runs can be stopped immediately if model behavior could cause real-world harm. According to the company, it hardened its evaluation environments following its own incidents, temporarily stopped multiple evaluations that give models internet access, and implemented real-time monitoring for web search evaluations that require internet access.
FBI warns that open-source AI penetration testing tool ARTEX has been weaponized
行业动态
The FBI warned that attackers have used the open-source AI penetration-testing framework ARTEX to automate intrusions since August 2026, with some obtaining elevated cloud privileges in under 3 hours. Personal information belonging to approximately 67,000 people was stolen from a South Korean financial institution.
Multiple attackers have been using the open-source AI penetration-testing framework ARTEX to automate intrusions since August 2026, targeting organizations in finance, government, defense, energy and other sectors, according to an FBI alert. The FBI said some intrusions took less than 3 hours from target identification to obtaining elevated cloud privileges, and personal information belonging to approximately 67,000 people was stolen from a South Korean financial institution. Attackers also used stolen ConnectWise cloud backup credentials to attempt logins to more than 2,250 Microsoft 365 tenants. Developer Autumn-27 released ARTEX in July 2026 and has since taken the project down and announced a move to closed source. As of early November 2026, researchers were still finding more than 1,600 publicly accessible control interfaces, while a commercial version, ARTEX PRO, was also circulating.
Cursor discloses security incident: leaked keys led to partial data exposure
行业动态
Following a two-week investigation, Cursor disclosed a security incident involving data it stored and shared, credentials it issued, and systems it operated. A malicious, unauthorized copy of the Claude Code extension also appeared in its extension marketplace; the company said the vast majority of customers were unaffected.
Cursor published information about the impact of a security incident following a two-week investigation and said it had contacted every affected customer directly with details specific to them. According to the company, the vast majority of customers were unaffected, and the incident involved data it stored and shared, credentials it issued, and systems it operated. Cursor had previously disclosed that it was investigating the incident and a malicious, unauthorized copy of Anthropic’s Claude Code extension that appeared in the Cursor VS Code extension marketplace. The company said it was addressing the impact and strengthening security measures based on its findings, and would update the announcement if anything important changed.
