Every source. One signal. The day in artificial intelligence, distilled into plain English.
Transmission 013Tuesday, 23 June 2026
Tuesday 23 June 2026 brings a dense mix of security warnings, model competition, workforce disruption and infrastructure manoeuvring. Intelligence agencies from five nations warn that catastrophic AI-enabled attacks on governments and critical infrastructure may be only months away, while OpenAI responds with a cybersecurity-focused model release. Elsewhere, Oracle's mass redundancies, Google's talent drain and a striking small-model efficiency result from researchers underline how quickly the competitive landscape is shifting.
Audio edition
Listen to today's transmission
—:——
Cyber threat and AI security
Signal 8/10
Five Eyes issues rare warning as OpenAI launches GPT-5.5-Cyber and security vulnerabilities multiply
A rare joint statement from the Five Eyes intelligence alliance — the United States, United Kingdom, Canada, Australia and New Zealand — warns that AI models capable of devastating attacks on governments and businesses may be ready within months. OpenAI responded by expanding its Daybreak programme with the full release of GPT-5.5 (Generative Pre-trained Transformer 5.5)-Cyber, aimed at helping defenders identify and patch vulnerabilities in open-source software through a 'Patch the Planet' initiative. Separately, security researchers clarified that Anthropic's Mythos model did not autonomously breach National Security Agency classified systems; the widely reported incident involved controlled red-team exercises rather than an actual intrusion. Microsoft patched a code-execution flaw in AutoGen Studio, its multi-agent development framework, and Simon Willison published analysis framing prompt injection as a form of role confusion rather than a purely technical exploit.
Small models punch above their weight as GLM 5.2 and a 3B-parameter reasoning model draw attention
China's GLM (General Language Model) 5.2 is generating strong user enthusiasm, with practitioners praising its performance relative to its size in a dedicated podcast discussion. On Hacker News, VibeThinker — a 3-billion-parameter model trained with a novel combination of supervised fine-tuning and group relative policy optimisation (GRPO) — reportedly outperforms Anthropic's Claude Opus 4.5 on reasoning benchmarks, a notable result given the parameter count differential. Sakana AI, a Japanese laboratory, launched Fugu and Fugu Ultra, orchestration models that route tasks dynamically across a pool of frontier large language models (LLMs) and reportedly lead most coding, reasoning and agentic benchmarks. Reports also surfaced that a Japanese AI system beat Claude 5 on certain benchmarks, though details remain limited. The throughline across all four items is that efficiency-focused training approaches and intelligent routing are closing the gap between smaller and flagship models.
Autonomous agents proliferate from coding environments to enterprise workflows as the 'loopy' era begins
TechCrunch describes a new phase of agentic artificial intelligence it calls 'loopy': swarms of agents authorised to work continuously in the background without human checkpoints, taking AI a significant step beyond single-task automation. xAI introduced '/goal' in Grok Build, a long-running autonomous execution mode that plans, executes a checklist and self-verifies until a goal is complete. AWS Lambda now offers micro virtual machines for isolated execution of user- and AI-generated code, lowering the barrier to safe agent deployment. Google made its Interactions application programming interface (API) the default interface for Gemini models and agents, retiring the old generateContent API and signalling that new agent features will ship exclusively through the new interface. Amazon is testing its upgraded Alexa+ assistant in India with Hindi-language support, marking a significant expansion of conversational AI into a major non-English market.
AI in the workplace: surveillance, trust and labour
Signal 8/10
Meta halts employee keystroke programme after internal data leak as AI-driven job cuts spread
Meta paused an internal initiative that collected workers' keystroke data to train AI models after a security failure left the data accessible across the company's internal systems; the exposure prompted significant internal concern and external coverage. The incident coincides with a broader wave of AI-attributed redundancies: Oracle announced cuts of 21,000 jobs as it pivots further towards AI, General Motors installed robots at its flagship electric vehicle factory after laying off 1,300 workers, and TechCrunch maintains a running list of major 2026 tech layoffs where employers cited AI as a factor. In a counterpoint, Anthropic announced it will pay workers US$85,000 to learn AI skills, illustrating how companies are simultaneously shedding some roles and investing in others. The Guardian also explored the first successful use of an AI law firm in an English court, with the presiding barrister noting that human advocacy remained central to the trial.
Alphabet shares slide on talent exodus, Groq confirms $650 million raise and AI political spending intensifies
Alphabet's stock recorded its worst single-day fall in more than a year after two senior AI researchers departed for OpenAI and Anthropic respectively, with Bloomberg and CNBC reporting that investors are rattled by the talent drain; these are reported market reactions, not guaranteed indicators of underlying performance. AI chipmaker Groq confirmed a $650 million funding round and announced it is rebuilding its executive team following a $20 billion non-acquisition deal with Nvidia, positioning itself as a neocloud provider. The Wall Street Journal reported on how Sam Altman's personal investment portfolio benefits from his ties to OpenAI, raising governance questions. XTX Markets, the quantitative trading firm that profited from an early Anthropic stake, is reportedly seeking further AI investments. Meanwhile, AI-focused political action committees are spending heavily in the 2026 United States midterm elections, with a single Manhattan congressional primary absorbing a large share of reported outlays.
Nvidia proposes hotter-running, lower-water data centres as the AI infrastructure market is projected to reach $810 billion by 2033
Nvidia has unveiled a new data centre cooling architecture for its Rubin platform that runs at higher temperatures in order to reduce water consumption inside facilities; TechCrunch cautions that this addresses only in-facility water use and does nothing to reduce the water embedded in fossil-fuel power generation that feeds these sites. A market analysis cited by Yahoo Finance projects the global AI data centre market to reach $810.6 billion by 2033 as enterprise investment accelerates — this is a projected figure from a market research report and should be treated as illustrative, not predictive. Micron and Anthropic signed an AI infrastructure supply agreement in which Micron is investing in Anthropic's Series H funding round and will supply memory components for Claude's infrastructure; The Decoder notes that critics regard such circular investment structures as potentially inflating valuations. Samsung is rolling out OpenAI tools to its workforce in what is described as one of OpenAI's largest enterprise deals, signalling continued hyperscaler and conglomerate commitment to AI infrastructure spend.
Claude Code's extended thinking output may not reflect genuine reasoning, new analysis finds
A widely read Hacker News post (498 points) argues that the text appearing in Claude Code's 'Extended Thinking' mode is not an authentic record of the model's internal reasoning process, raising questions about what users are actually observing when they read a model's apparent chain of thought. This connects to a broader concern in AI safety and evaluation research: if the visible reasoning trace is post-hoc rationalisation rather than a real computational path, benchmarks and audits built around inspecting that trace may be unreliable. The Moebius project, also highlighted on Hacker News, offers a practical counterpoint — a 0.2-billion-parameter image inpainting model achieving performance comparable to 10-billion-parameter models, demonstrating that careful architecture choices rather than scale can deliver strong results, and Simon Willison documented porting Moebius to run in a web browser using Claude Code, providing a concrete replicability test for the model's claims.
From AI boyfriends to AI filmmaking: the cultural frontier expands as Google DeepMind backs A24
The Guardian published a long-read by a self-described AI sceptic who tested an AI romantic companion, finding the experience alternately unsettling and revealing about human attachment — a piece that prompted zero news-feed engagement but reflects growing public curiosity about AI social roles. At the institutional end of the creative spectrum, Google DeepMind committed $75 million to a partnership with independent film studio A24 to develop AI filmmaking tools, a significant signal of intent to enter Hollywood production pipelines. The Verge reported on AI-generated virtual staging in real-estate listings, describing how AI imagery is creating impossible-looking homes that distort renters' expectations before they view properties. Together, these items mark a broadening of AI's cultural footprint beyond productivity into intimacy, entertainment and housing.
Port a small open-source AI model to run entirely in your browser using Claude Code
Simon Willison documented how he used Claude Code to port the Moebius image inpainting model — a 0.2-billion-parameter model with strong performance — so that it runs directly in a web browser with no server required. The technique is practical for any developer who wants to let users run a lightweight model locally without infrastructure costs or data-privacy concerns.
Find a small, well-documented open-source model (Moebius or similar) with published weights in a compatible format such as ONNX (Open Neural Network Exchange) or safetensors.
Open Claude Code and describe your goal: converting the model to run in the browser via WebAssembly or a JavaScript inference library such as ONNX Runtime Web or Transformers.js.
Use Claude Code's Extended Thinking or agentic mode to iteratively generate, test and debug the JavaScript conversion code, asking it to explain each step.
Test the resulting page locally in a browser, checking memory consumption and inference speed against the original Python implementation.
Publish the browser-runnable demo (for example via GitHub Pages) and compare outputs with the server-side version to validate parity.