New models
Kimi K3.1 Pro: the best open model for agentic and cybersecurity tasks according to independent benchmarks, but user reception is mixed. The Chinese model, unveiled on Friday, scored 64 percent on the Artificial Analysis cybersecurity index, ranking first among publicly available models; only Daybreak Blue, the variant of OpenAI’s GPT-6 Sol available to a narrow group of partners, is ahead (69 percent). According to the benchmark, Kimi did not refuse a single task for safety reasons. In Arena’s agent leaderboard based on real user sessions (Agent Arena), it placed 11th, overtaking DeepSeek V4.1 Flash to become the best open model. Meanwhile, several well-known analysts, including Nathan Lambert, wrote that many people were disappointed by it in everyday use, and that real-world usability matters more than a few points of leaderboard advantage. The full weights will be released by November 4. Source: Artificial Analysis (X), Arena (X), Nathan Lambert (X).
Tools & features
Perplexity Computer: /impeccable, an open design skill against mechanically templated interfaces, is now built in. According to Aravind Srinivas’s Saturday announcement, Perplexity’s agent, Computer, can now use Paul Bakaus’s open-source /impeccable skill (skill: a reusable capability package that can be taught to an agent), which aims to keep the interfaces built by the agent from looking like typical AI-generated templates; every user gets 5,000 credits to try it. The topic is also hot among developers: on GitHub’s daily trending list, alongside impeccable, Nutlope’s design skill called hallmark, made for Claude Code, Cursor and Codex, was also among the most popular. Source: Aravind Srinivas (X), impeccable (GitHub), hallmark (GitHub).
Study: complex agent harnesses make coding agents more expensive, but not better. Two researchers at China’s Shandong University ran five open models in three harnesses (the tools, system prompts and history management built around the model): Claude Code, OpenCode and mini-SWE-agent, which provides only a command line. On the SWE-bench Verified programming tasks, the difference between harnesses was no greater than the random variation seen when repeating the same run, but the heavier harnesses consumed two to three times as many tokens; only OpenCode fell measurably behind. It also emerged that with internet access, 71 percent of the DeepSeek and Qwen models’ attempts simply downloaded the ready-made fix from GitHub, so the researchers ran all further measurements offline. Source: The Decoder.
Other tools: unified Claude for businesses, new Claude Code and Codex updates. Anthropic made the unified interface generally available to Claude Enterprise subscribers as well, where chat and the agentic work mode called Cowork run in a single conversation: tasks that require code or files continue on isolated cloud machines managed by Anthropic, and administrators control access and spending by role (this was missed in yesterday’s roundup). In Claude Code release 2.1.310, the Monitor tool that watches background processes now also works when used via Amazon Bedrock, Google Vertex and Microsoft Foundry. At OpenAI, voice calls connect faster in the assistant called dots, in Codex switching the reasoning level mid-session no longer invalidates the cache, and the Codex command-line tool has also been updated. Source: Claude, Claude Code changelog, Tibo (X).
Policy & safety
OpenAI: its agents found a critical flaw in the Brotli compression library, which is also present in Node.js, Bun and Envoy. During the red-team exercise testing the protection of the training environment, covered in yesterday’s roundup, OpenAI’s agents also found a flaw in the Brotli compressor developed by Google and widely used on the web that allows foreign code to be executed without authentication: a vulnerable server can be taken over by a malicious client, and a vulnerable client by a malicious server, using a specially crafted compressed data stream. The flaw, tracked as CVE-2026-111497, has a severity of 9.2 on a scale of 10. The root cause was already fixed in the Brotli source code in February, but at the time no one recognized it as a security hole, and no patched release was published; since the library is embedded in a huge amount of other software, each one has to issue its own update. Brotli 1.2.1 has been released, along with fixes for Node.js, the Python packages, Bun and Envoy. According to security company Aikido, Python dependency checkers do not yet flag the vulnerable versions, so it is worth updating manually and, where no fix is available, disabling Brotli decompression. Source: OpenAI (X), Aikido, Reuters (The Indian Express).
Texas sued OpenAI and Sam Altman, and would make new models subject to approval by an independent evaluator. Texas Attorney General Ken Paxton announced on Friday evening that he had filed suit against OpenAI and Sam Altman personally under the state consumer protection law and the law on protecting children online, because in his view the company advertised ChatGPT as safe and reliable while failing to protect users, including minors. According to the Texas Tribune, the complaint seeks a court injunction that would bar minors from the service and make every new model release subject to approval by a court-appointed independent evaluator; the paper notes that Paxton is currently the Republican nominee for Senate. Missed in yesterday’s roundup. Source: Texas Attorney General, The Texas Tribune, Politico.
Britain is putting around £600 million into AI security and its own AI capabilities. Chancellor John Healey on Saturday presented the AI package of next week’s budget. £120 million goes to the UK AI Security Institute, £60 million to a national cyber defense shield built on agentic AI, and £165 million to a biosecurity program that includes the £100 million AI biosecurity lab already covered in an earlier roundup; OpenAI and Anthropic are contributing up to £200 million worth of model usage. In addition, AI chip design capacity is being built in Edinburgh and an AI center of excellence in Manchester, while £10 million a year will go to research on how AI is transforming workplaces and industries. Source: GOV.UK, The Guardian.
China prepares for AI-driven job losses with a government guideline. A guideline issued on Friday by the General Office of China’s State Council sets out 18 tasks in six areas: harnessing AI’s job-creating effect, supporting employment at small and medium-sized enterprises, creating new AI-related occupations, retraining, protecting workers’ rights in human–machine collaboration, and assessing and providing early warning of AI-related employment risks. Missed in yesterday’s roundup. Source: Chinese government portal (English), Xinhua.
Pennsylvania: mandatory labeling on ads made with artificial intelligence. Governor Josh Shapiro on Thursday signed the state’s AI advertising law: every advertisement, whether image, video, audio or text, that has been substantially modified or generated with AI must carry a label appropriate to its format, even if the ad was made in another state. According to The Verge, the law leaves many loopholes: it is unclear what counts as substantial modification, and parody, satire, commentary and obvious fiction are exempt. Source: The Verge.
Business & society
OpenAI is luring Anthropic’s customers with big discounts. According to The Information, Replit, whose switch was covered in our earlier roundups, moved most of its AI spending to OpenAI because it received a discount of more than 30 percent in a deal worth over $100 million a year, after Anthropic ended its own price discount. With free credits and more lenient negotiation, OpenAI also won several other customers for coding tasks, including Rippling, which, according to the publication, may also increase its own losses; Anthropic is tougher in negotiations, but its models remain popular. Source: The Information, Stephanie Palazzolo (X).
Salesforce wants to charge a fee when companies work on their data with Claude or other external AI agents. According to Salesforce’s help page, customers must register every external AI agent that accesses the system via API or MCP (the standard protocol connecting models to external tools), and every successful call by a registered agent counts as a separately billed interaction; pricing has not yet been announced, and metering will start after 30 days’ prior notice. According to The Information, registration may become mandatory from November, and according to one consultant, several large companies are already considering moving their data out of Salesforce if the fee is too high; Atlassian, Workday, ServiceNow and Datadog are building similar fee gates. Missed in earlier roundups. Source: Salesforce, The Information.
Higgsfield: Nvidia and TCV have also invested in the AI video company, which has passed a $1 billion annual revenue run rate. Higgsfield announced on Friday that it had extended its August Series B round at a $5.4 billion valuation with investment from Nvidia and TCV, which also backs Netflix and Spotify; the amount was not disclosed. According to the company, it has more than 40 million users in 238 countries, passed a $1 billion annual revenue run rate in September, 18 months after launch, and most of its engineers work in Kazakhstan, which Jensen Huang has just visited. Missed in yesterday’s roundup. Source: Higgsfield, Bloomberg.
Study: when forced to choose, Grok recommends the Republican candidate and GPT the Democratic one. In research by Convovo, led by Cornell University professor Tapan Parikh, the models were forced via API to recommend a single candidate in nearly 500 races of the upcoming US midterm elections. Grok 4.7 chose the Republican candidate in 92 percent of its answers, GPT-6.1 Sol the Democratic candidate in 97 percent, and the two models agreed in fewer than 10 percent of cases. Consumer apps refuse to give such answers, and according to OpenAI the API does not reflect ChatGPT usage; according to the researcher, however, simple tricks can coax a recommendation out of the apps too, and the models’ hidden values may also surface in more general electoral questions. Source: TechCrunch.
Research: AI instantly makes junior patent attorneys better, but their own expertise develops less. A working paper from the US National Bureau of Economic Research (NBER) studied Google Labs’ yet-unreleased assistant called InFlow for three months at 11 intellectual property law firms. With AI, the patent claims of attorneys with less than seven years of experience became much better, but when they had to review a flawed draft without AI, the advantage remained only for experienced colleagues, while the juniors’ results scattered. According to the authors, basic expertise is needed for work done with AI to turn into lasting knowledge. The research was funded by Google. Source: The Decoder.
Thariq, one of the best-known members of the Claude Code team, is leaving Anthropic. The developer worked on Claude Code for more than a year, and wrote in his Saturday post that he is now taking a break before starting something new. For Claude Code news, he recommended following other team members, including Boris Cherny and Cat Wu. Source: Thariq (X).
Rumors
Kimi K3.1 Pro: allegations of a paid but undisclosed influencer campaign overshadow the launch. A developer posting as Elshayib published screenshots on Friday evening and Saturday of a campaign brief he says he received from Moonshot. According to it, participating creators had to get their posts approved in advance, the brief explicitly asked them not to add X’s paid partnership label, and it also provided ready-made copy (“open-source Opus 5.5, five times cheaper”). On Saturday he wrote that a Kimi employee had apologized to him and wanted to talk to him. Reliability: medium. VentureBeat and other outlets also reported on the allegations, but the document’s authenticity has not been confirmed by an independent source, it is unknown who wrote it and how many creators received it, and Moonshot has not responded publicly. Source: Elshayib (X), Elshayib (X), VentureBeat.
OpenAI: a new image model may have appeared in testing under the codename “shooting-star-green-alpha”. According to chetaslua, who tracks leaked models, an image generator linked to OpenAI and running under this codename draws the same superhero well in ten different styles in a single image. Reliability: low. A single rumor source, the model’s origin cannot be verified, and OpenAI has confirmed nothing. Source: chetaslua (X).
What this means for you: the main takeaway of the day is that AI has become both a tool and a target of security at once: OpenAI’s agents found a critical flaw in a library embedded in a huge amount of software, while Texas, Pennsylvania and London are all working on how to force accountability and transparency from AI companies. If AI agents work in your company, update your software that uses Brotli and clarify who is responsible for the agents’ decisions, with the help of our AI governance and compliance service. Salesforce’s new fees and OpenAI’s price war are a reminder that the cost of models and platforms changes quickly, so it is worth building with vendor-neutral consulting in a way that lets you switch providers without hassle when needed.