# Horizon AI Briefing — 2026-11-04

## Top stories
- **Postman Silently Auto-Installs AI Plugins and Attempts to Share All API Data** (https://github.com/postmanlabs/postman-app-support/issues/14370) — Postman is reportedly auto-installing AI plugins without user consent and attempting to share API data, raising serious IP and security concerns for enterprise developers. This mirrors a broader pattern of developer tools quietly expanding data collection under the guise of AI features, and professionals should audit their toolchain settings immediately.
- **OpenAI Allegedly Re-Enabling Model Training on Codex Chats Without User Consent** (https://news.ycombinator.com/item?id=50352757) — Multiple users report that OpenAI's 'Improve the model for everyone' setting in Codex was silently re-enabled after they had explicitly turned it off, potentially exposing proprietary code and prompts to training pipelines. This is a critical IP risk for enterprise teams using AI coding assistants and warrants an immediate check of data-sharing settings at chatgpt.com/settings/data-controls.
- **Google Ties K-12 Search Access to AI Adoption, Drawing 'Blackmail' Criticism** (https://bsky.app/profile/newrepublic.com/post/3mzgpasmv5o2o) — Google has disabled Gemini and generative AI features by default on K-12 accounts, but crucially, schools that want to block AI cannot disable AI-infused Search while keeping standard search available — effectively forcing schools to accept AI tools to retain full search functionality. Critics and independent reporters are characterizing this as coercive bundling, with significant implications for EdTech procurement and AI regulation debates.
- **US Tech Giants' Hidden AI Debt Swells 90% to $3.2 Trillion in Six Months** (https://asia.nikkei.com/business/technology/artificial-intelligence/us-tech-giants-hidden-ai-debt-swells-90-to-3.2tn-in-6-months) — Independent reporting finds that off-balance-sheet AI-related financial obligations at major US tech companies have ballooned 90% to $3.2 trillion in just six months, raising systemic financial risk concerns. Professionals tracking AI infrastructure investment should note this as a potential leading indicator of a correction or repricing event.
- **Anthropic Launches FlashProve: Cryptographic Proof of AI Model Identity** (https://www.itmedia.co.jp/aiplus/article/2611/04/2000002863/) — Anthropic announced FlashProve, a zkSNARK-based system that cryptographically verifies an AI response originated from a specific model, preventing spoofing or model substitution. Released as a research prototype with accompanying code, this addresses a growing enterprise concern about supply-chain integrity in AI deployments. [self-reported; source linked]
- **OpenAI Partners with Cathedral, a Firm Founded by Former DOGE Personnel** (https://bsky.app/profile/hagenblix.bsky.social/post/3mzgmyodmck27) — OpenAI is reportedly partnering with Cathedral, a company formed by former DOGE operatives including Marko Elez, who drew attention for publicly racist statements. The partnership raises governance and reputational questions for OpenAI as it approaches a more public-facing growth phase.
- **Stanford Study: 44% of Widely-Cited LLM Research Claims Have Already 'Expired'** (https://the-decoder.com/rapid-llm-development-is-outrunning-the-research-meant-to-explain-it-study-finds/) — Stanford researchers tested 50 high-profile LLM findings against 20 current models and found that 44% are already invalidated and another 34% have weakened significantly, with an average claim lifespan of just 15.7 months. This has major implications for how enterprises and policymakers should weight AI research when making procurement or regulatory decisions.
- **SemiAnalysis GDPx Benchmark Shows Top AI Agents Score Only 12% on Real Enterprise Tasks** (https://www.aitimes.com/news/articleView.html?idxno=216936) — SemiAnalysis released GDPx, a benchmark testing frontier AI models on actual professional tasks spanning semiconductors, cybersecurity, biology, finance, law, and software engineering — with the best models achieving only 12%. The stark gap between lab benchmarks and enterprise reality is a critical data point for organizations evaluating AI agent ROI.
- **HBM Chip Smuggling to China: Samsung and SK Hynix Memory Hidden Inside Low-End Processors** (https://www.aitimes.com/news/articleView.html?idxno=216978) — Independent reporting reveals Chinese entities are concealing high-bandwidth memory chips from Samsung and SK Hynix inside low-performance server processors to circumvent US export controls, then extracting and supplying the HBM to AI chipmakers like Huawei. This exposes a significant loophole in the current export control regime with direct implications for AI chip supply chain security.
- **Anthropic IPO Filing Reveals $1.7B Quarterly Payroll, ~$107K Per Employee Monthly** (https://www.techspot.com/news/114428-anthropic-ipo-filing-reveals-17-billion-quarterly-personnel.html) — Anthropic's IPO filing discloses a quarterly payroll of $1.7 billion — approximately $107,000 per employee per month — offering a rare window into the cost structure of a frontier AI lab at scale. This comes alongside news of departures from the Claude Code team, including product lead Cat Wu, raising questions about pre-IPO talent retention. [company-reported figures; source linked]

## Where accounts diverge
- Account 0 states the acquisition value is approximately $11.5 billion, while Account 1 states it is $115 billion. [quotes translated by Horizon]
  - itmedia.co.jp: "approximately $11.5 billion" (https://www.itmedia.co.jp/news/article/2611/03/2000002830/)
  - aitimes.com: "$115 billion" (https://www.aitimes.com/news/articleView.html?idxno=216952)

## Emerging signals
- **AI Developer Tools Quietly Expanding Data Collection — A Pattern, Not an Anomaly** — Both Postman and OpenAI Codex are independently accused of silently enabling data-sharing features that expose user IP and API data, suggesting a systemic trend of AI toolmakers treating user data as a resource to be harvested by default. Enterprises need explicit data governance policies for every AI-integrated dev tool in their stack.
- **Weight-Space Programming: AI Directly Editing Another AI's Weights** (https://www.aitimes.com/news/articleView.html?idxno=216971) — MIT researchers report that one AI model can directly modify another's neural network weights — changing just 0.016% of parameters — to achieve what they self-report as over 6x improvement in math scores, without any retraining. Per their own benchmark claims (no linked paper), this 'weight-space programming' approach could redefine how models are specialized and updated post-deployment.
- **Enterprise AI Model Hijacking Attacks Surging, With Losses Exceeding $1M** (https://www.etnews.com/20261104000243) — Google Cloud Threat Intelligence and Mandiant researchers document a spike in LLM account hijacking attacks on enterprise AI deployments, with single-day damages exceeding $100,000 and cumulative losses reaching $1 million in documented cases. As agentic AI expands enterprise attack surfaces, this signals that AI security is moving from theoretical to materially costly.
- **AI Political Influence Spending Concealed Behind Non-AI Ad Campaigns** (https://bsky.app/profile/bellingcat.com/post/3mzfx5667n223) — Independent reporting finds AI industry leaders and companies are spending millions to influence US midterm elections and AI regulation, but their political ads make no mention of AI — a stealth lobbying pattern with significant implications for AI governance and transparency.
- **Pre-IPO Talent Exodus at Frontier AI Labs** (https://bsky.app/profile/carnage4life.bsky.social/post/3mzh5n4ovqk2j) — Anthropic is seeing a notable departure wave ahead of its IPO: two Claude Code engineers left in October, and now product lead Cat Wu has also departed. This emerging pattern of pre-IPO exits at frontier labs warrants monitoring as a leading indicator of internal culture or strategic tensions.

## New entrants
- **FlashProve** (tool) — Anthropic's zkSNARK-based cryptographic system that proves an AI response originated from a specific model, released as a research prototype with open code. Addresses model identity and supply-chain integrity concerns.
- **GDPx** (framework) — SemiAnalysis's new AI agent benchmark evaluating frontier models on real enterprise professional tasks across six domains; top models score only 12%, exposing the gap between lab and real-world performance.
- **Copilot Fusion** (tool) — GitHub's officially launched feature that automatically selects the optimal AI model for each coding task to balance performance and cost, announced at GitHub Universe 2026. Capability claims are self-reported with no linked independent evaluation.
- **EXAONE Nexus** (tool) — LG AI Research's AI agent for automated legal risk detection in AI training data, claiming 1,000-fold cost reduction in copyright compliance verification. Capability claims are self-reported.
- **Google Antigravity** (tool) — Google's agentic cheminformatics platform that accelerates drug discovery workflows, enabling compound search and novel molecule design. Capabilities are self-reported.

## Biggest movers this week
- **Apple** (company) — 23 mentions this week, ↑14 vs the prior week
- **Sam Altman** (person) — 20 mentions this week, ↑9 vs the prior week
- **Dan Driscoll** (person) — 7 mentions this week, ↑7 vs the prior week
- **Qwen 3.6 35B** (model) — 6 mentions this week, ↑5 vs the prior week
- **Z.ai** (company) — 6 mentions this week, ↑5 vs the prior week
- **Alphabet** (company) — 5 mentions this week, ↑5 vs the prior week

## China & East-Asia AI
- **Lenovo Kaitian Unveils AI Host With Zhaoxin KX-7000 CPU and Houmo LQ50 NPU for Local 35B Models** (https://pandaily.com/lenovo-kaitian-ai-host-zhaoxin-kx-7000-houmo-lq50-npu-160-tops-local-35b-qwen) — Pandaily
- **Baidu’s new AI simulator gives robotaxis a crash course in rare road hazards** (https://www.scmp.com/tech/big-tech/article/3373173/baidus-new-ai-simulator-gives-robotaxis-crash-course-rare-road-hazards) — South China Morning Post
- **👧🏻Chinese AI Has a Persona** (https://www.recodechinaai.com/p/chinese-ai-has-a-persona) — Recode China AI
- **DeepSeek Expands AI Agent Capabilities with Natural Language—Automating Plugin Creation** (https://www.aitimes.com/news/articleView.html?idxno=216958) — AI Times
- **Why Chinese AI stars like Z.ai, MiniMax struggle to match Anthropic’s profit margin** (https://www.scmp.com/tech/big-tech/article/3373124/why-chinese-ai-stars-zai-minimax-struggle-match-anthropics-profit-margin) — South China Morning Post

## Korea AI
- **Hidden HBM Chips in Cheap Processors...Samsung and SK Hynix Smuggling Methods to China Exposed** (https://www.aitimes.com/news/articleView.html?idxno=216978) — AI Times
- **OpenAI: ChatGPT Search Volume Reaches 20% of Google's Total, Weekly Users Surpass 1.2 Billion** (https://www.aitimes.com/news/articleView.html?idxno=216977) — AI Times
- **AI Directly Modifies Another AI's Weights Without Retraining, Achieving 6x Improvement in Math Scores** (https://www.aitimes.com/news/articleView.html?idxno=216971) — AI Times
- **Government-Led 'Humanoid Data Factory' Launches with 12 Robots in Operation** (https://www.aitimes.com/news/articleView.html?idxno=216973) — AI Times
- **Warning: Enterprise AI Model Hijacking Attacks Spike; $100K+ Damage Cases Reported** (https://www.etnews.com/20261104000243) — ETNews

## Japan AI
- **GitHub Officially Releases Copilot Fusion, Selecting Optimal AI Models by Task for Both High Performance and Low Cost** (https://www.itmedia.co.jp/news/article/2611/04/2000002869/) — ITmedia
- **Anthropic Releases FlashProve: Cryptographic Proof That AI Responses Come from a Specified Model** (https://www.itmedia.co.jp/aiplus/article/2611/04/2000002863/) — ITmedia
- **Claude Plan Rename Reduces Capacity by Half? Official Anthropic Denies It — 'Max 20x Weekly Limits Were Originally 10x Pro'** (https://www.itmedia.co.jp/aiplus/article/2611/04/2000002862/) — ITmedia
- **Apple Opens Refreshed 'Siri AI' to All Developers; App Integration Enables Reservations, Orders, and More** (https://www.itmedia.co.jp/news/article/2611/04/2000002855/) — ITmedia
- **3D CAD Data and AI Support Shipbuilding Design, Streamlining Design Information Verification Work** (https://monoist.itmedia.co.jp/mn/articles/2611/04/news020.html) — ITmedia

## Europe (EU) AI
- **OpenAI says GPT-6 Astra designed a hardware device from scratch in under a week** (https://the-decoder.com/openai-says-gpt-6-astra-designed-a-hardware-device-from-scratch-in-under-a-week/) — The Decoder
- **Rapid LLM development is outrunning the research meant to explain it, study finds** (https://the-decoder.com/rapid-llm-development-is-outrunning-the-research-meant-to-explain-it-study-finds/) — The Decoder
- **EU Court of Justice Advocate General: AI Chatbots Can Violate Publishers' Copyright Under Certain Circumstances** (https://t3n.de/news/ki-chatbots-eugh-urheberrechte-1771465) — t3n
- **AI agents take over at SETI's Allen Telescope Array, scanning for extraterrestrial intelligence** (https://the-decoder.com/ai-agents-take-over-at-setis-allen-telescope-array-scanning-for-extraterrestrial-intelligence/) — The Decoder
- **OpenAI hires ex-Army Secretary, gaining bipartisan reach just as midterms loom** (https://the-decoder.com/openai-hires-ex-army-secretary-gaining-bipartisan-reach-just-as-midterms-loom/) — The Decoder

---
Source: Horizon (https://horizon.alchemylab.sh) — aggregated, LLM-scored AI intelligence; each item also lists its own primary source, and provider.citation credits both.