← Dailies

Daily · 2026-11-03 00:00 UTC

Anthropic flags near-RSI capability; Marvell buys Lumentum for $115B

Anthropic's IPO filing warns of recursive self-improvement, Marvell buys Lumentum for $115 billion, Goldman Sachs pulls AI coding tools, and a 50-agent Minecraft swarm builds a bureaucracy.

Source ↗

🎧 News Digest — 2026-11-03 00:00 UTC 10 items · ~3 min listen

1. Anthropic's IPO filing warns it may be near recursive self-improvement — In a revised S-1A filed Monday, Anthropic added risk-factor language saying it has already used AI to develop models whose capability leaps it did not identify at the time, and that internal estimates indicate it may soon meet its own definition of "recursive self-improvement." It points to its Responsible Scaling Policy as the control on that risk. TechCrunch · Nov 2
2. Marvell buys Lumentum for $115B in an all-stock deal — Enterprise value of $115 billion including debt, equity value about $80 billion, with 4.5 Marvell shares for each Lumentum share. The deal is expected to close in the first half of calendar 2028, pending shareholder and regulatory approval. Nvidia invested $2 billion in Lumentum and $2 billion in Marvell earlier this year. TechCrunch · Nov 2
3. Goldman Sachs pulls AI coding tools from traders and bankers — Per Bloomberg reporting relayed by The Decoder, internal guidelines from October limit Claude Code and GitHub Copilot to engineers. Traders and bankers had been using them to monitor flows and draft content. A Goldman spokesperson cited "productivity and risk management" and said the experimentation phase is over. The Decoder · Nov 2
4. A 50-agent Minecraft swarm invented a chicken handoff protocol — Y Combinator Labs and Olam Labs put 50 GPT-6.1 Sol agents into a Minecraft survival world for three hours with one instruction: build a thriving civilisation. They exchanged about 5,900 messages, suffered 22 deaths on the first night, then turned extremely risk-averse and bureaucratic — including a written item-tossing ban after a cooked chicken was eaten by the wrong agent, and an 11-minute debate over a "Sole Fertilizer Officer." In nine 45-minute Bed Wars rounds against Claude Opus 5.5 agents, the GPT team won 6-3, at one point downloading the game's source code to work around a bug. The Decoder · Nov 2
5. UK cyber agency: AI-powered malware likely to hit critical infrastructure by 2028 — The National Cyber Security Centre's annual review says it is "highly likely that between now and 2028, AI-powered malware will have been used in a ransomware attack on UK CNI," and warns Russia and China will use AI "routinely" in attacks. Chief executive Richard Horne is expected to warn of "an unprecedented rise in the number of cyber incidents" in the coming months. The Guardian · Nov 3
6. 29 NHS trusts listed as "realising benefits" from Palantir say it isn't live — A Guardian investigation built on more than 100 freedom-of-information requests found that of the trusts NHS England publicly lists as benefiting from the £330m federated data platform, 29 said the system was not live or they were not using any of its applications. One senior health official called the gap "very misleading"; NHS England says nearly 140 trusts have reported benefits. The Guardian · Nov 2
7. Google says a manufacturing partner built a backdoor into 1.3 million devices — Shenzhen-based Videostrong turned Android TV set-top boxes, karaoke machines, telehealth devices and digital signage into a botnet used for ad fraud, data exfiltration and residential proxy traffic. Google says it dismantled the backdoor, cut ties, and banned the manufacturer from certified Android devices; disrupting the affiliate command-and-control domains may sever access to at least 10 million more compromised devices. The Verge · Nov 2
8. OpenAI pitches data centres as "Grid Protectors" — A new pilot at an upcoming data centre in Effingham County, Georgia promises to free up to 1,000 megawatts of power during extreme weather or emergencies — enough, OpenAI says, to power air conditioners in around 700,000 homes. The announcement lands as local opposition to data centres spreads and several states move to block cost pass-through to residents. The Verge · Nov 2
9. Deep Principle releases Hydrogen-1.0, a 753B open-weight science model — Built on GLM-5.3, the model is aimed at scientific problem solving and research assistance: explaining derivations, organising responses to open-ended research questions, and supporting work in scientific computing, reaction-network exploration, NMR structure analysis and experimental planning. Weights are on Hugging Face, with NVFP4 builds for NVIDIA Blackwell and MXFP4 for AMD Instinct MI350X. Hugging Face · Nov 2
10. Google documents "LLMjacking" bills of $100K a day — A Mandiant and Google Threat Intelligence Group report covers cases where attackers hijacked enterprise LLM access and ran up more than $100,000 in fraudulent cloud charges in a single day; at some organisations the cumulative bill reached $1 million. Entry points included stolen service-account keys and GitHub personal access tokens, LLM endpoints exposed to the internet such as Ollama servers, and a user-ID rebind flaw in LiteLLM (CVE-2026-40217). SiliconANGLE · Nov 2

🧪 Try This Next

• 🔧 Pull Deep Principle's Hydrogen-1.0 from Hugging Face and test it on a derivation from your own field — the NVFP4 and MXFP4 builds are ready for Blackwell and MI350X hardware.
• 🎥 Read the Y Combinator Labs and Olam Labs swarm writeup before your next multi-agent build; the risk-aversion spiral is the part that transfers to production agents.
• 📝 Audit which cloud keys and model endpoints your team leaves exposed — every LLMjacking case in the report started with a stolen token or an open endpoint.

Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.