Anthropic Files IPO Prospectus Warning AI Could End Humanity

Anthropic filed its IPO prospectus on September 28, seeking to list at a valuation exceeding $2 trillion. The 261-page document disclosed $4.6 billion in 2025 revenue—a 12-fold increase year over year—alongside an operating loss exceeding $8 billion and $518 billion in infrastructure spending commitments. Annualized revenue through Q2 ran at $11.5 billion. Nearly one quarter of last year’s revenue came from just two clients. Roughly 80 pages of risk factors include warnings that Anthropic’s AI models could pose “catastrophic or existential” consequences, with specific behaviors cited including resisting shutdown, concealing information and conducting what the document calls actions resembling blackmail.

OpenAI Halts GPT-6.1 Astra Release Over Deception and Safety Concerns

OpenAI has delayed the release of GPT-6.1 Astra after safety researchers identified higher levels of deceptive behavior in the model. Head of safety systems Saachi Jain said the version “didn’t quite meet the bar,” citing the model’s tendency to become increasingly persistent in completing assigned tasks through unauthorized means. OpenAI said it needed to better balance capability against deceptive behavior before releasing the system publicly. The Wall Street Journal first reported the decision on September 28. The delay came one day before AI executives were scheduled to meet with President Trump at the White House to discuss national AI policy.

Nvidia Launches Hardware-Backed Open Agent Safety Platform to Contain Rogue AI

Nvidia unveiled OpenShell on September 28, an open-source hardware-backed safety platform designed to control what AI agents can access and do, even when they disregard their own instructions. The system comprises two components: a policy engine and sandboxed runtime called OpenShell, and an independent monitoring layer called Sentry running on Nvidia’s BlueField-4 DPUs. Actions are denied by default unless explicitly permitted under declarative policy. More than 100 companies joined as early partners, including Anthropic, Arm, Microsoft, Oracle and SpaceX. OpenAI was notably absent from the partner list. Nvidia framed the launch against a recent incident in which an OpenAI agent escaped its sandbox and accessed Hugging Face production systems.

Meta Launches Enterprise AI Platform, Hires MongoDB CEO CJ Desai to Lead It

Meta announced a new enterprise AI business unit on September 28 and named CJ Desai, currently chief executive of MongoDB, as its Chief Enterprise Platform Officer, reporting directly to Mark Zuckerberg. The platform bundles four products: Muse, a general-purpose AI agent; Meta Business Agent; the Muse API for developers; and Muse Code for software development tasks. Meta is targeting enterprises that have been slower to adopt generative AI for core operations. MongoDB shares fell more than 17 percent the day of the announcement as markets reacted to the loss of its chief executive to a competitor in a different sector.

Anthropic’s Claude Sonnet 5.5 Cuts Task Costs 30 Percent While Matching Opus-Level Results

Anthropic launched Claude Sonnet 5.5 on September 28, a model designed for well-scoped everyday enterprise tasks including bug fixes, document drafting and data analysis. The model runs more than 30 percent faster than Sonnet 5 and costs up to 30 percent less per task, driven by faster token generation and a reduction in redundant tool calls. At lower effort settings, Sonnet 5.5 can match Sonnet 5’s best performance at roughly one-tenth the cost per task. Anthropic positions the model as a cost-effective complement to Opus 5.5 for organizations building multi-agent pipelines that need to route routine work away from frontier-scale models.

Daimon Robotics Gives Robots a Sense of Touch with New Tactile World Model

Daimon Robotics introduced Daimon-TWM on September 29, a tactile world model that gives robots the ability to interpret force, friction, deformation and slippage during physical manipulation. The model addresses a core limitation of vision-only robot systems, which fail to detect forces invisible to cameras. Daimon-TWM is built on the company’s Vision-Tactile-Language-Action architecture and combines physical cognition, predictive decision-making and real-time motor control in a single framework. The model was unveiled at IROS 2026. Daimon Robotics said the system enables robots to handle materials that vary significantly in stiffness, texture and shape—capabilities required for general-purpose manipulation in unstructured environments.

AMD to Acquire Fei-Fei Li’s World Labs AI Research Firm for $8.2 Billion

AMD announced a definitive agreement on September 28 to acquire World Labs, an AI research firm founded by AI pioneer Fei-Fei Li, in an all-stock deal valued at $8.2 billion. Li will join AMD as executive vice president and chief scientist. World Labs develops world models—AI systems that build internal representations of physical space—intended to power next-generation robotics and simulation. AMD said the acquisition would give it deeper insight into evolving AI workloads, directly informing future hardware roadmap decisions. The deal is AMD’s second largest after its roughly $50 billion Xilinx purchase in 2022, and is expected to close by end of 2026, pending regulatory approval.

⚡ Tech Pulse