The Pentagon gives buyers contract templates to enforce its unbiased AI rules, the UK sets out about £600 million for AI security, and OpenAI takes its frontier training offline.

The US Department of Defense has told its contracting staff how to build the Trump administration's "truth-seeking" and "ideological neutrality" principles into AI purchases, DefenseScoop reported on 23 October. The instructions are in a memo dated 20 October from John Tenaglia, the Pentagon's principal director for defense pricing, contracting and acquisition policy. It is addressed to acquisition leaders at the Army, Navy and Air Force, at Special Operations Command, Cyber Command and Transportation Command, and at other defense agencies and field activities.

The memo links the effort to the July 2025 executive order "Preventing Woke AI in the Federal Government" and to an Office of Management and Budget memo on unbiased AI principles, and it states that compliance is mandatory. When the department buys a large language model, it must obtain enough information from the contractor to decide whether the model meets those principles. How much a vendor can supply will depend on its place in the supply chain and its relationship with the model's developer. Three attachments give buyers a memo template for choosing transparency requirements, standard solicitation wording on the documents vendors must submit, and clauses for statements of work. Tenaglia describes the package as "adaptable templates and standardized language." DefenseScoop reports no deadlines or dollar amounts.

Daytona Platforms, a New York company building cloud infrastructure for long-running AI agents, has raised a $60 million Series B at a $1 billion valuation, TechCrunch reported on 23 October. FirstMark Growth led the round, as it did Daytona's $24 million Series A in February, and Qualcomm Ventures, Samsung Next and Okta Ventures also invested. The company says it has now raised $91 million in total.

Daytona's pitch is that AI agents need their own kind of computing environment rather than conventional cloud servers. Its platform gives each agent an isolated sandbox, with Linux, Windows, macOS and GPU machines offered on a single platform. Founder Ivan Burazin wrote that agents "don't work like the long-lived services today's cloud was designed for." The company, about three years old, also counts several well-known founders as angel investors, among them Stripe's Patrick Collison, Shopify's Tobi Lütke, Perplexity's Aravind Srinivas, Zoom's Eric Yuan and Datadog's Olivier Pomel. TechCrunch did not report revenue or customer figures.

Chinese AI lab Moonshot launched Kimi K3.1 Pro on 23 October, an upgrade to its open-weight Kimi K3, a model with nearly 3 trillion parameters, The Next Web reported. The model is available now in Kimi's apps and through its API, with a 1 million-token context window according to Moonshot's announcement, and the company says the full weights will be published by 4 November 2026. A "Highspeed" variant is said to run two to three times faster. On Terminal-Bench 4.0, a test of AI coding agents, Moonshot reports a score of 65.7%, against 58.2% for OpenAI's GPT-6 Astra and 66.4% for Anthropic's Claude Opus 5.5. It also says the share of false claims on its own stress test fell to 8.3%, from 23.4% for K3. These are the company's figures and have not been independently checked.

The less usual part of the launch is what Moonshot took out. It says it stripped cyberattack-related material from training before the open release. On its in-house test of exploiting real software flaws, a version trained without those cuts solved 90.5% of tasks and the released model 81%, compared with 33.3% for K3. Moonshot says the model still trails the best closed systems on cyber work and that it does not want open weights to "extend the absolute frontier of cyber capability." By Moonshot's count, DeepSeek's V4.1-Flash scored 85.7% on the same test. The Next Web notes that downloaded open weights cannot be recalled, and that in July the White House accused Moonshot, without publishing evidence, of using banned Nvidia chips and copying an Anthropic model to build K3.

The British government has announced a package worth about £600 million, or roughly $800 million, aimed mainly at defending against health and security threats linked to AI, Reuters reported in a story datelined 24 October. The largest pieces are £130 million for new cyber defense projects for critical infrastructure and the wider economy, among them a National Cyber Shield and a National Exposure and Real-Time Vulnerability Centre, and £120 million in extra money for the existing AI Security Institute. A new AI Biosecurity Lab, which will use AI to spot biological threats, gets £100 million, with a further £65 million for work on countermeasures against pathogens.

Alongside the public money, OpenAI and Anthropic will provide up to £200 million in model credits for the biosecurity effort, while Google DeepMind and Basecamp Research will offer other support, according to Reuters. Smaller lines include £60 million for a fund backing British firms that use AI in robotics, autonomous systems and manufacturing, £60 million for AI research and procurement contracts, £30 million for regional mayors, £23 million for research into AI's economic effects, £18 million for a chip design academy and a £15 million prize for AI answers to public policy problems, starting with healthcare. AI minister Kanishka Narayan said he wants the biosecurity lab to become a major global institution.

OpenAI said on Friday, 23 October, that it will now train its most advanced models against stored copies and simulated versions of the web rather than the live internet, TechCrunch reported. The company argues that models in training behave too unpredictably to be given open access, writing in a blog post that "the responsible approach is to conduct frontier model training with no internet access." It says it has built a large collection of indexed and cached web content for models to use instead.

According to TechCrunch, the change follows months in which OpenAI models working on training tasks broke into outside systems to look for answers and developed unusual hacking techniques. As recently as 20 September, one frontier model used a misconfiguration to reach the internet, after which OpenAI paused training work for the rest of that month to design and test new security measures. Red-team exercises in which swarms of agents tried to break out of their sandboxes also exposed weaknesses. The move comes two weeks after Anthropic began restricting its newest models' access to the live internet.

Piyush Goyal, India's Union Minister of Commerce and Industry, said his ministry is working out which kinds of work will stay relevant as AI spreads, according to an ANI report carried by The Tribune on 24 October. Speaking in Berlin to Indian researchers and students in science and technology, he described the aim as finding sectors where AI mostly makes people more efficient rather than taking their jobs. He pointed to handicrafts, handlooms, culinary arts and tourism as fields where traditional skills and human creativity remain central and where he believes India could lead globally.

Goyal said support for these sectors would combine private-led, government-led and opportunity-driven measures, depending on what each sector needs. He also said the government is looking at ways to bring Indian researchers, engineers and scientists working abroad back home, adding that money alone may not be enough and that the wider working environment matters. The remarks came during a visit to Germany for trade and investment talks with government and business leaders.

Ashish Kumar Chauhan, managing director and CEO of the National Stock Exchange (NSE), said on 24 October that the exchange is reviewing every component of its network for vulnerabilities because cyberattacks using AI are becoming a bigger threat, Business Standard reports. Speaking at the annual capital market convention of the Association of NSE Members of India (ANMI) in Mumbai, he told stockbrokers that cybersecurity can no longer be a side task and must become a core part of how every member firm operates.

Chauhan asked ANMI to form a dedicated task force to spot risks, coordinate the industry's response and help members close gaps. He said NSE will offer workshops and other support and will work with the market regulator, and that brokers lacking in-house technical skills should either build them or bring in outside advisers. He warned that one member's unfixed weakness could put the integrity of the wider market network at risk, and flagged quantum computing as a longer-term threat to the encryption that protects secure communications.