The Hal Times

The United Kingdom has announced the creation of a National AI Biosecurity Laboratory, backed by a £100 million budget. The initiative is positioned as a safeguard against the misuse of artificial intelligence in biological research, and it is the first government‑funded centre of its kind. The announcement was made jointly by the Department for Science, Innovation and Technology and the Office for Life Sciences, underscoring the cross‑departmental importance of the project.

The laboratory will be staffed by a consortium of universities, national health agencies and private‑sector partners. OpenAI and Anthropic have been granted technical credits for their expertise in large language models and safety‑oriented AI design. Their involvement is intended to accelerate the development of containment protocols and rapid‑response tools that can detect and neutralise AI‑generated bio‑threats.

What sets this lab apart is its explicit focus on the intersection of AI and biological security, rather than general AI safety. It will host dedicated compute clusters isolated from the public internet, and it will run simulated pathogen‑design scenarios to test defensive measures. The budget will fund both hardware and a year‑long programme of ethical‑risk assessments, with an eye toward future regulatory frameworks.

The announcement leaves several questions unanswered. The timeline for operational readiness is not specified, and the governance structure for oversight remains vague. It is also unclear how the credits awarded to OpenAI and Anthropic will be quantified or audited. Observers will watch for further details on staffing, data handling policies and the mechanisms by which the lab will coordinate with international biosecurity bodies. (NewsCord)

OpenAI Cuts Live‑Internet Training for Frontier Models

OpenAI has announced that it will cease training its frontier models on the live internet, citing significant risks associated with uncontrolled data ingestion. The company says the decision follows internal assessments that the approach could expose models to disinformation, privacy breaches and regulatory hazards. Training will now be confined to curated, offline datasets that undergo rigorous vetting before use.

In a complementary technical brief, OpenAI described enhancements to its training environment, including isolated compute pods, high‑compute red‑team exercises and hardened containment layers. These measures aim to prevent inadvertent leakage of model outputs and to strengthen evaluation protocols. The changes mark a shift toward greater operational security, but the impact on model performance and development timelines remains to be seen. (TechCrunch; OpenAI)

OpenAI Partners With Poland’s NASK on AI Safety

OpenAI has entered a partnership with Poland’s National Academic Computer Centre (NASK) to advance research on AI safety and security. The collaboration will focus on developing technical safeguards, improving resilience against adversarial attacks and sharing best practices for responsible deployment. Both organisations intend to co‑author papers and contribute to open‑source toolkits that can be adopted by the broader research community.

The partnership also includes joint workshops aimed at building a more robust digital ecosystem in Central Europe. While the agreement outlines shared objectives, the specific milestones, funding allocations and deliverable timelines have not been disclosed. Stakeholders will watch for forthcoming publications that reveal the concrete outputs of the joint effort. (OpenAI)

Anthropic IPO Valuation Guidance from Damodaran

As Anthropic prepares for an initial public offering, finance professor Aswath Damodaran has offered a framework for investors to assess the company’s value. He highlights the importance of scrutinising the firm’s compute spend, model performance metrics and revenue streams from enterprise licences. Damodaran cautions that the high‑cost nature of large‑scale AI research can amplify valuation volatility.

The analysis suggests that traditional multiples may under‑represent the long‑term risk profile of AI‑centric firms. While the guidance provides a starting point, uncertainties around future model adoption, regulatory environments and competitive dynamics remain unresolved, leaving investors to hedge against a range of possible outcomes. (The Economic Times)

From the Laboratories

•  AI‑focused compute spending reaches $750 billion, with Anthropic leading the charge, while OpenAI trims its own budget and Nvidia offers a safety net for developers. (AI: Reset to Zero)

•  The AI giants' speculative bets on compute and talent could expose public pensions to heightened financial risk. (The Independent)

•  Hugging Face introduces LitPivot, a tool for dynamically contextualising research literature to surface emerging ideas. (Hugging Face)

•  Developers can now query whether their models run on AWS Trainium directly from the Hugging Face hub. (Hugging Face)

•  Turning database grading into reinforcement‑learning rewards enables post‑training agents to improve via the ThinkingBox framework. (Hugging Face)

•  OpenAI details its second‑phase storage scaling that delivers twelve‑fold faster reads for over a billion ChatGPT users. (OpenAI)

Moonshot AI Release

Moonshot AI Announces Kimi K3.1 Pro Model

Moonshot AI has introduced its latest large language model, Kimi K3.1 Pro. The architecture is a mixture of experts design, utilising 104 billion active parameters out of a 2.8 trillion total parameter pool. The model supports a context window of one million tokens. It is currently available via the Kimi API, though it does not yet feature open weights. The company has committed to releasing the full model weights by 4 November 2026.

The firm has also introduced a high speed version of the model, which is reported to perform two to three times faster than the standard release. This version carries a price premium, costing twice as much as the standard model. Input token cache hits are excluded from this price increase and remain at the same rate. Pricing for the standard model is set at 2.00 dollars per million tokens for input and 10.00 dollars per million tokens for output.

The model has entered the market with a score of 55 on Artificial Analysis, securing a position on the Pareto line. Moonshot AI has also implemented new safety measures within this release. The developers have explicitly curbed the hacking capabilities of the model in response to concerns regarding the potential use of artificial intelligence in cyberattacks. (MarkTechPost; Simon Willison; Artificial Analysis)

Anthropic Launches Security Scanner For Open Source

Anthropic has launched a service called OSS Scanner aimed at open-source project maintainers. The tool is free and requires an opt-in from the maintainer to begin scanning. The service uses the company's language models, including Claude Mythos, to identify potential vulnerabilities within project code.

Security reports are generated automatically without human intervention. Maintainers receive these findings and retain control over which fixes they wish to apply to their projects. This service is designed to automate the discovery of security flaws in the open-source ecosystem. (InfoQ)

In Brief

•  Simon Willison has documented a tool using Claude Opus 5.5 High to perform image analysis and rendering with simulated brushstrokes. (Simon Willison)

•  Hugging Face has released a guide for developers to determine if their models are compatible with AWS Trainium hardware. (Hugging Face)

•  Microsoft researchers have published details on ThinkingBox, a system that uses database grading to provide rewards for training agents. (Hugging Face)

•  A new project called LitPivot provides researchers with methods for dynamic contextualization of academic literature. (Hugging Face)

Union Tech Policy

IATSE Appoints First Technology Intelligence Officer

Jillian Arnold has taken the role of labor and technology intelligence officer for IATSE. As the first official in this position, she aims to navigate the arrival of artificial intelligence within the film and television industry. Arnold states her intention to help reskill crew members as automated tools become more common on set.

The union seeks to ensure that workers have a voice during the creation of these new technologies rather than merely reacting to their deployment. By involving craftspeople in the development process, the union hopes to guide how software shapes the working day. Success depends on the cooperation of production studios.

The extent of this influence remains unknown. While the appointment signals a shift in strategy, the union must still negotiate the boundaries of machine labour against the traditional skills of its members. The objective is to maintain professional standards while adapting to the presence of new software in the arts. (Hollywood Reporter)

Lawsuits Challenge AI Training Methods

Reservoir Media has filed copyright infringement lawsuits against Suno and Anthropic. These legal actions highlight the tension between rightsholders and firms that develop artificial intelligence models using existing creative works. The litigation follows a trend of music companies seeking to establish clearer rules for the use of their assets in training data.

The outcome of these court cases will carry weight for the industry. If the courts find that the use of copyrighted material for training purposes constitutes infringement, AI companies may face higher costs and restricted access to data. For now, the legal proceedings move through the courts. (Music Ally)

Steam Next Fest Sees Rise In AI Usage

Generative AI usage has reached 36 percent at the latest Steam Next Fest. This figure represents a significant increase in the number of developers integrating automated tools into their game design and production workflows.

The rise in adoption reflects a broader push to shorten development cycles through the use of machine learning. Players and critics continue to monitor these releases to determine how such tools alter the creative output and polish of new titles. The impact of this shift on game quality remains a subject of ongoing debate among the gaming public.

In Brief

•  China has debuted its first fully AI-generated film.

•  Canva has released an image generation model trained on design templates rather than photographs. (PetaPixel)

•  Formula 1 fans suspect the sport is using artificial intelligence for recent camera transitions. (PetaPixel)