Amazon's Andy Jassy personally flagged the concerns that triggered Anthropic's ban
The Anthropic model suspension story deepened as WSJ reported that Amazon CEO Andy Jassy personally flagged security concerns to US officials that triggered the crackdown on Fable 5 and Mythos 5, while India began debating strategic independence from American AI providers and a coalition of state attorneys general opened a separate investigation into OpenAI. Zhipu AI released GLM 5.2 with a 1-million-token context window and MIT open weights arriving next week, and China’s government forced Meta to unwind its $2 billion Manus acquisition on national security grounds.
Regulatory & Policy #
Amazon CEO’s talks with U.S. officials triggered crackdown on Anthropic models #
WSJ / TechCrunch / Hacker News (697 points)
WSJ reported that Amazon CEO Andy Jassy informed government officials, including Treasury Secretary Scott Bessent, that Amazon researchers had used Anthropic’s Claude Fable 5 “to obtain information that could be used in cyberattacks” – conversations that directly triggered the Commerce Department’s export control directive covered in yesterday’s digest. Former AI czar David Sacks claimed a jailbreak vulnerability had been discovered, and when Anthropic refused to remove the model, the administration acted. The revelation that Amazon – Anthropic’s largest investor at $4 billion – initiated the government response adds a layer of complexity: the company that stands to benefit most from Anthropic’s success also triggered the most disruptive regulatory action against it.
OpenAI faces investigation from state attorneys general #
TechCrunch
A coalition of state attorneys general has opened an investigation into OpenAI, with New York’s AG serving a subpoena requesting documents on advertising practices, user engagement metrics, model behavior, consumer and health data handling, and treatment of minors and seniors. This follows lawsuits over copyright infringement and ChatGPT’s alleged connection to user suicides, as well as a separate suit from Florida’s AG claiming the company ignored safety warnings. The breadth of the subpoena – spanning advertising, health data, and vulnerable populations – suggests regulators are examining OpenAI as a consumer product company, not just an AI lab.
As Anthropic suspends access to new models, India debates its AI future #
TechCrunch
The Anthropic model suspension affected India just after the company announced a major partnership with Tata Consultancy Services, triggering a debate among Indian tech leaders about reducing dependence on American AI providers. The incident crystallized a geopolitical risk that was previously theoretical: “American AI models are bound to American geopolitics,” as one policy expert noted. Whether India can meaningfully accelerate domestic AI capabilities remains contested, but the political will to try has shifted from aspirational to urgent.
American Government Takes Down Claude #
Don’t Worry About the Vase (Zvi Mowshowitz)
Zvi’s policy analysis of the Fable/Mythos suspension identifies three structural problems: the implementation simultaneously restricts AI model access while the same administration relaxes chip export controls to China; demanding “no jailbreaks” is technically impossible since every LLM has adversarial vulnerabilities; and if the standard were applied consistently, it would halt all frontier model deployments industry-wide. The piece warns that foreign AI researchers may flee US labs and the action undermines American AI competitiveness through self-sabotage, though Zvi reserves final judgment pending whether the government’s demands prove to be narrow and reversible or a permanent new regime.
Model Releases #
GLM 5.2 released: Zhipu AI’s 1M-context coding-first model #
Zhipu AI / Hacker News (601 points)
Zhipu AI launched GLM 5.2, a 744-billion-parameter mixture-of-experts model with a usable 1-million-token context window and up to 131,072 output tokens per response – roughly five times the context capacity of its predecessor GLM-5.1. The model is positioned as coding-first with a dual thinking-effort system (High and Max modes) and is live across all GLM Coding Plan tiers. API access, the Z.ai chatbot, and MIT-licensed open weights are scheduled for the following week. Notably, Zhipu published no benchmark numbers at launch, so performance claims are entirely unverified – a pattern worth watching given the model’s aggressive positioning against frontier competitors.
Security #
The future of Siri, or: why private inference isn’t private enough #
Matthew Green / Lobsters
Cryptographer Matthew Green argues that Apple’s Private Cloud Compute, while technically sound for inference privacy, cannot protect AI agents that need to communicate externally to be useful. Agents that access private data must query search engines, send messages, and interact with third-party services – each interaction leaking information that cryptographic protections cannot cover. The analysis identifies three threat vectors: corporate monetization of agent queries, remote prompt injection attacks to exfiltrate private data, and government surveillance through agent-accessible channels. Green’s conclusion is stark: “only legal and political safeguards” can address these risks, since the useful capabilities of AI agents are architecturally incompatible with meaningful privacy guarantees.
Research & Papers #
SFT Drives Gemini’s Safety Properties #
Google DeepMind / AI Alignment Forum
Google DeepMind’s Language Model Interpretability team reports a surprising finding: Gemini’s safety characteristics stem primarily from supervised fine-tuning combined with pretraining, not from reinforcement learning or later post-training stages. The team applied SFT using Gemini’s training mixture to pre-training-only versions of Gemini 3.1 Pro and 3 Flash, then compared them against production models across safety benchmarks – the results were “remarkably similar.” This identifies SFT as a higher-leverage intervention point for safety than the RL-based approaches that have received more research attention, and suggests that the safety properties of deployed models may be more attributable to training data curation than to the alignment techniques applied afterward.
Funding & Business #
Meta reportedly moves to unwind $2B Manus deal after Beijing’s demand #
TechCrunch
Meta is dismantling its $2 billion acquisition of Chinese AI startup Manus after Beijing ordered the deal reversed on national security grounds approximately two months ago. Meta has already severed Manus from its internal systems and halted data sharing, while Manus co-founders are exploring raising approximately $1 billion from outside investors to reclaim the startup – potentially leading to a Chinese joint venture and Hong Kong listing. The action mirrors the US government’s own intervention against Anthropic this week: both Washington and Beijing are asserting sovereign control over AI assets they consider strategically sensitive, regardless of corporate structure or offshore incorporation.
Developer Tools #
AI coding at home without going broke #
Stephen Bochinski / Hacker News (299 points)
A practical breakdown of the cost landscape for individual developers using AI coding tools, recommending a hybrid approach: expensive frontier models for specification writing and hard reasoning tasks, cheaper open-source models via API providers like OpenRouter for routine implementation. The core insight is that spec-driven development – investing upfront in detailed specifications before generating code – minimizes expensive token usage while maintaining output quality. The post estimates a balanced strategy can deliver significant AI-assisted development output at roughly $1,000/month, providing a concrete cost benchmark for practitioners evaluating whether to self-host, subscribe, or use APIs.
Infrastructure #
RTX 5080 and RTX 3090 setup: 80+ tok/s on Qwen 3.6 27B Q8 #
imil.net / Hacker News (254 points)
A detailed configuration guide for running a 27-billion-parameter model at 80+ tokens per second using consumer-grade GPUs with speculative decoding, achieving a 230K context window on hardware costing a fraction of data center infrastructure. The key findings are operational: BIOS configuration (UEFI mode, Above 4G Decoding, ReSize BAR) is critical for dual-GPU functionality, and the standard nvidia-open driver outperforms alternatives for heterogeneous GPU setups. The result demonstrates that local inference on current open-weight models has crossed a practical threshold where consumer hardware can serve near-real-time responses for individual developer workflows.
Other #
KPMG pulls report on AI usage due to apparent hallucinations #
TechCrunch
KPMG withdrew its October 2025 report on “agentic AI” after GPTZero identified numerous fabricated claims – organizations including UBS, the NHS, Swiss Federal Railways, and Transport for London reported that the report’s statements about their AI usage were “either untrue or misleading.” The firm apparently used AI to write a report about AI, producing hallucinated case studies that no human reviewer caught before publication. Beyond the irony, the incident demonstrates that established institutional brands are not immune to AI-generated misinformation, and that professional services firms face reputational risk when they deploy AI without rigorous fact-checking workflows – exactly the kind of validation that their clients pay them to provide.
Threads to Watch #
AI governance is being tested from every direction simultaneously. US export controls on Anthropic, state AG investigations into OpenAI, Beijing forcing Meta to unwind its Manus acquisition, and India debating AI sovereignty – all within 48 hours – demonstrate that governments are no longer treating AI regulation as a future concern. The unanswered question is whether these interventions are coordinated enough to produce coherent policy or whether they will fragment the global AI market into incompatible regulatory zones.
Open-weight models are the hedge against regulatory fragility. GLM 5.2’s MIT-licensed weights, coding-first positioning, and 1M-token context window arrive at exactly the moment when closed-model access has proven fragile – Anthropic’s most capable models can be pulled with a government letter, but weights on disk cannot be recalled. The RTX 5080/3090 local inference setup achieving 80 tok/s on a 27B model shows the complementary hardware trend: running open models locally is crossing from hobbyist experiment to practical developer workflow.
Trust in AI is under strain at every layer. KPMG’s hallucinated report demonstrates that AI-generated content can slip through professional review processes, Google DeepMind’s SFT finding suggests alignment may rely on simpler mechanisms than widely assumed, and Matthew Green’s analysis argues that the privacy protections AI agents actually need are beyond what technology alone can provide. The common thread is a gap between what AI systems are assumed to do and what they verifiably do – a gap that regulators, researchers, and users are all discovering independently.