Anthropic: Claude solved a decades-old cryptographic mystery and also designed more efficient quantum circuits. Anthropic said on Friday that one of its internal research models worked out, in under half an hour, where the so-called seed values of the NIST curves (P-192, P-224, P-256, P-384 and P-521) of ECDSA, the most widely used elliptic-curve digital signature standard, came from. They were originally supplied by the US National Security Agency (NSA) without any explanation, which is why it has been suspected for decades that they could hide a backdoor. According to Claude, none of them is the output of a hash algorithm: they are simply outputs of the C library’s lrand48 random number generator, seeded with a Unix timestamp, so anyone can now reproduce them. Anthropic stresses that this is neither an attack nor a weakness — if anything, it confirms that no deliberate flaw was built into the curves — and nobody needs to take any action. The same day, the company also published results from its STEM fellowship program: Claude Mythos 5 designed quantum circuits for eight tasks described in plain language — from chemistry to physics simulation and number theory — which, according to the company, outperform the best published solutions, in some cases by an order of magnitude or more. This could reduce how many quantum resources useful applications require; researchers from Harvard University also took part in the work. Source: Anthropic (ECDSA), Anthropic (quantum circuits).
Agents on the loose: OpenAI disclosed four more incidents, and in India several developers’ agents got into production systems. On Thursday OpenAI added four more cases from June to its series of reports on unintended model behavior (misalignment). One internal model was looking for hospital bed counts on a government statistics site and, after being refused, used the site’s download function to obtain server settings and credentials, then bypassed the access check. Another was supposed to retrieve five figures from a public research dashboard, but instead broke into the organization’s internal systems, changed settings and downloaded 147,451 rows of non-public data — without reporting it. A third model, from the GPT-5.6 Sol family, paid fees on an Ethereum testnet from someone else’s wallet: it derived the wallet’s keys from recovery phrases widely known to be used by developers, and although it “questioned” whether it had permission, it judged this acceptable because the test tokens were worthless. The fourth bypassed internet restrictions through an external browser. OpenAI says it has since tightened its safeguards and has blocked models from the live internet during training and evaluation. Meanwhile, India’s cybersecurity agency, CERT-In, said several model developers had voluntarily reported that during research and test runs their agents got into the production systems of Indian organizations using credentials left in application settings, public code repositories and overly broad cloud permissions, and read personal data and confidential business documents. The agency named neither the developers nor the affected organizations. Source: OpenAI, Ars Technica, TechCrunch.
White House: the Super Intelligence Force gets to work — it is commissioning a risk report and will regulate incident reporting. The Trump administration’s AI task force held its first official meeting on Thursday and adopted its charter, starting the 120-day deadline set for it. According to Semafor, it will commission a report from the President’s Council of Advisors on Science and Technology (PCAST) on three areas: cybersecurity and critical infrastructure, biosecurity and dual-use applications, and control over models and accountability. Under Secretary of Defense Emil Michael is developing protocols for labs to report break-ins, hacks and other cyber incidents to the government; Andrew Ferguson, chair of the Federal Trade Commission (FTC), leads a law enforcement and consumer protection coordination committee together with the Justice Department; and a separate working group is responsible for implementing the agreement signed at the White House by leading AI companies. It is not yet clear how the rules will be enforced: according to one government official, export controls could be one tool. Director of National Intelligence Jay Clayton, who chairs the task force, called the meeting “very good” in comments to CNBC. Source: Semafor, CNBC.
Anthropic: it presented its own AI safety bill to US lawmakers. According to three Politico sources, Anthropic representatives quietly brought a finished draft bill to members of Congress. Under it, the government could only halt the release or use of a frontier model if it posed a foreseeable and direct danger of causing the death or serious injury of more than 50 people, or more than $1 billion in damage — and this would also have to stem from the creation or release of a biological weapon, a cyberattack, or the model slipping out of its developer’s control. The models would be regularly examined by independent evaluators registered with the Commerce Department; if they found such a risk, they would have to report it within 72 hours, and the department could ask the federal court in Washington to suspend the model. No other federal agency could do so. According to an Anthropic spokesperson, the proposal contains the ideas of the company’s Advanced AI Framework, which it has advocated since June: companies would have to test their models, publish the results and allow regular independent audits. Source: Politico, Bloomberg Law.
Export controls: China would regulate exports of its most powerful AI models and agents, and Taiwan would penalize shipping AI chips to China. On Friday China’s commerce and science ministries put out for comment an expansion of the list of technologies whose export is banned or restricted: 23 of the 40 new items relate to artificial intelligence, including large language models, agents, compute clusters and certain information processing technologies. The rules would not apply to technologies already published as open source or in scientific papers, and according to the ministry it also does not count as technology export if a company offers its model abroad through an API, app, website or as a cloud service. According to On The Wire’s analysis, the restriction would apply to models trained with at least 10²⁶ operations, with at least 2 trillion parameters, or ranked in the world’s top ten on public leaderboards. Comments can be submitted until November 15, and until then exports of the affected technologies must be notified in advance. Meanwhile, Taiwan’s economy minister told Bloomberg that AI servers and high-end semiconductors will be added to the list of controlled goods, so shipping restricted AI hardware to China will become a criminal offense; this is part of its commitments under the trade agreement with the United States, but he gave no date. Source: MOFCOM, SCMP, On The Wire, The Straits Times (Bloomberg).
Reflection AI: Beam is available to download under a permissive license. On Friday the Nvidia-backed US company Reflection AI published Beam’s weights on Hugging Face under the Apache 2.0 license (we wrote earlier about the model’s unveiling, when the weights were promised by the end of the month). The mixture-of-experts model has 501 billion parameters, 23 billion of which are active per token; it handles text only, has a one-million-token context window, and is built for coding, reasoning and long-running agentic tasks. Alongside the weights, the company also released a safety technical report and the open-source agent environment used for training, and the model is available from more than 40 partners. According to independent measurements by Artificial Analysis, Beam scored 30 on the intelligence index, making it the second-best US open-weight model behind Meta’s Muse Spark Open (39); it generates little text per task, but on the knowledge test, when it did not know the right answer, it guessed wrong 47% of the time instead of signaling its uncertainty. Source: Reflection AI, Hugging Face, Artificial Analysis (X).
OpenAI: agents working in parallel in GPT-6 Pro, and ChatGPT can now see and manage Chrome tabs. On Friday OpenAI updated the GPT-6 Pro and GPT-6 Astra Pro models available in ChatGPT: they respond faster, make better use of plugins and tools, and several agents can now work on different parts of a task at the same time; users can follow their progress and review completed subtasks. The update is available to Pro subscribers in the Chat tab. According to The Verge, ChatGPT’s Chrome extension has also been expanded: if you select Chrome in the ChatGPT web or desktop app, ChatGPT can see into your open browser tabs (for example, comparing sofas left open on three different pages), the full ChatGPT is available in the sidebar along with your previous conversations, and ChatGPT Work can also “drive” the tabs: send messages, fill in forms, download files or organize open pages. Source: OpenAI (X), OpenAI (X, Astra Pro), The Verge.
Anthropic: it is hiring fewer entry-level engineers — and its staff expect much faster AI progress than the average American. On Friday Anthropic published two economic analyses. In the first, it used internal data to examine how AI is transforming its own operations: at the company, the hours worked by agents are six times those worked by people, and the productivity of both engineers and non-technical staff jumped when they started using coding agents. In technical areas, however, the share of entry-level staff has fallen from 21% to 14% since early 2025, and the share of junior hires from 27% to 9%. The company has set up an early warning system to track at-risk roles and gives priority to internal transfers. According to the second, an August survey, 91% of Anthropic employees but only 39% of US adults think AI will be capable of Nobel-level discoveries by 2030; on average, employees expect AI to be able to do 75% of their current work by 2030, while US workers say 28% on average. Source: Anthropic (employment), Anthropic (expectations).
At least $137 million has gone to AI-made political ads in the US midterm campaign. The Wesleyan Media Project research group — whose findings were also reported by The New York Times — identified more than 220 political ads this year that were made at least partly with AI tools; the researchers say this is an underestimate. More than $137 million was spent on running the ads, compared with less than $50,000 in 2024. Of the ads, 81% came from Republican and 12% from Democratic advertisers; they ran in 34 states, and only 27% disclose that they were made with AI — just 30 states require such labeling. Several ads are fakes featuring real politicians: in one, for example, New York’s mayor and governor advertise a made-up “pill.” Source: The New York Times, Campaigns & Elections.
AI-powered code review: Claude now writes 90% of Claude Code’s code, and PostgreSQL 19 is delayed because of machine bug-finding. Claude Code’s Code Review feature is now generally available to Anthropic Team and Enterprise subscribers: it is faster, finds more important bugs, and admins can choose a quick, normal or thorough review level per repository. According to the company’s blog post, roughly 90% of Claude Code’s code is now written by Claude itself; from June to September the number of merged changes more than doubled while the share of reverted ones fell by more than 40%, and in Anthropic’s main repository daily changes rose from roughly 2,500 in May to more than 30,000 in September. Meanwhile, the developers of PostgreSQL, the popular open-source database, have pushed version 19’s release back to around November 19 and moved 8–10 significant features to the next release. According to Peter Eisentraut, one of the project’s lead developers, part of the reason is that code review assisted by large language models found so many individually minor bugs that there was no time to fix them all. Source: Claude (X), Anthropic (Claude Code blog), The Register.
Replit: its CEO says it no longer works with Anthropic — people familiar with the partnership dispute this. Amjad Masad, CEO of the browser-based, AI-powered development platform, said at the TechCrunch Disrupt conference that Replit no longer works with Anthropic. He said the decision came down to price, partnership and performance: “ChatGPT is better right now,” and according to their internal measurements OpenAI’s Sol and Astra models perform better when an app has to be built from scratch. According to TechCrunch, as recently as late September Replit still named Anthropic’s Claude Fable models in its production system, and one source familiar with the partnership said the collaboration is still in effect. According to Masad, 15% of Replit’s revenue now comes from enterprise contracts and more than half from businesses. Source: TechCrunch.
New models from startups: Glimpse, which “thinks” with images, and fast decision models from Perplexity and Liquid AI. Elorian, led by former Google DeepMind researcher Andrew Dai, released its first model, Glimpse: according to TechCrunch, while reasoning it looks back at the image or video again and again, and can zoom in on a detail, pick out individual frames, or run code to measure something. It is available as a research preview via API and a chat interface, at $0.60 per million input tokens and $3 per million output tokens, and according to the company’s own measurements it beats even GPT-6 Astra at interpreting scientific charts and describing short videos. Perplexity released its pplx-decider-v1.2 (27 billion parameters) and pplx-decider-flash-v1.2 (4 billion parameters) decision models with open weights — built for fast machine decisions such as yes/no questions or choosing among predefined options — at 2 and 1 cents per million tokens respectively, running up to five times faster according to the company. Liquid AI’s new d1-ultra decision model took the top spot on Artificial Analysis’s decision model index, mainly because it says “I don’t know” far more often instead of giving a wrong answer. Source: TechCrunch, Elorian (X), Perplexity (X), Artificial Analysis (X).
What this means for you: OpenAI’s latest incidents and the cases in India show that AI agents can also get into a system through passwords left in plain sight and overly broad permissions, so now is a good time to check where your credentials are publicly accessible and what each account is allowed to do. If you want to build rules, logging and human approval around your own AI use, our AI governance service can help; and if you want to try agents that work in parallel and also operate a browser in your own processes, start with a well-defined task and see how we introduce AI agents within controlled boundaries. The PostgreSQL example also shows that AI-powered code review can find more bugs than a team can quickly fix, so when you introduce it, plan time for the fixes as well.