Top Stories
Meta pauses Muse share-by-link after researchers expose thousands of private chats
The Verge reports that security researchers at Claroc say they enumerated roughly 6,800 Muse share links by mining public DNS records, then viewed hundreds of them and read what was supposed to be private. “Muse builds each share’s web address from the share’s title, a short phrase describing the conversation, then a random ID,” Claroc wrote. “The hostname usually gave away what a page held before we opened it, and since the link carries that same title and ID, a hostname pulled from DNS rebuilds the link.” Among the contents were a marketing company’s revenue review, a small-business owner’s loan application, a workplace’s full-year staff calendar, and notes for a job candidate’s call with an AI lab.
Meta has paused the share-by-link feature while it reviews the report. “People have the ability to share specific Muse artifacts, which creates a page that can be accessed by those with the link,” Muse spokesperson Tracy Clayton said in an emailed statement to The Verge. “However, we’ve paused this functionality while we review this report and consider changes to how these links can be discovered.” The failure mode is unusually clean: an enumeratable URL space that any researcher with passive DNS access can walk, which puts this in the same privacy category as prior share-by-link mishaps on ChatGPT, Claude, and Grok, but with a more reproducible leak path than the Google-indexing cases.
US orders federal civilian agencies to remove DeepSeek from their systems
A new Department of Homeland Security order published in the Federal Register and reported by TechCrunch directs federal civilian agencies to stop buying DeepSeek products and to remove existing ones from government IT systems, including third-party products and services that incorporate DeepSeek technology. The order extends a December 2025 law that covered intelligence and national-security systems into the wider federal civilian estate, citing concerns that sensitive information from US government users “could potentially be accessible to Chinese authorities under Chinese law.” Agencies can request exceptions on a case-by-case basis.
China’s foreign ministry pushed back. Foreign Ministry spokesperson Mao Ning said Beijing “opposes ‘overstretching the concept of national security and politicizing technology and trade issues,’” and added, “We will resolutely safeguard the legitimate rights and interests of Chinese enterprises.” The order lands alongside separate US allegations that DeepSeek, Alibaba, and Moonshot AI used large-scale distillation to pull information out of American AI systems: a September CISA report named DeepSeek and five other Chinese labs, OpenAI told lawmakers it had seen suspected DeepSeek attempts, and Anthropic said it identified unauthorized attempts by DeepSeek and other Chinese labs to collect Claude outputs. It also lands against a Bloomberg report that China is considering export restrictions on 23 AI-related technologies, including chip designs and certain advanced AI models. Practical takeaway for users: nothing in the order touches consumer phones or laptops, but expect vendor and supply-chain questions to ripple through state-government and federal-contractor procurement over the next few months.
Anthropic pitches an AI safety bill that could pull a model offline if 50+ people could be killed
TechCrunch outlines a draft bill Anthropic has been circulating that would order a model shut down once credible evidence shows a path to more than 50 deaths or more than $1 billion in property damage, with the trigger set by an independent evaluator registered with the US Commerce Department and confirmed by a DC judge. The bill defines catastrophic risk as arising from biological weapon creation or release, a model evading developer controls, or a major cyberattack, and it requires evaluators to notify Commerce within 72 hours.
A spokesperson told Politico, via TechCrunch, that the proposal “represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June.” Independent evaluators could be paid by AI labs, provided “payment is not contingent on the evaluations’ results.” The proposal also forbids federal agencies from independently halting distribution, leaving the courts as the gate. A frontier lab publishing this kind of binding-threshold legislation, tied to its own IPO runway, is itself the story worth tracking, especially given the $45 billion Nscale compute talks below.
”Reverse-Engineer Anything” turns decompiling into an AI-augmented wild west
404 Media covers a GitHub project with more than 342,000 stars that lets AI agents decompile, clone, and modify proprietary software at will, including smart TVs, smart picture frames, baby monitors, Lululemon mirrors, guitar pedals, Adobe Creative Cloud, Microsoft Office, and hundreds of video games. The piece quotes Mike Nolan: “All I am hearing from software engineers is that software-as-a-service is dying as a business model. It’s becoming so easy to recreate the functionality that you want [with AI.]”
A user on X posted: “Vibe coded my parents’ TV this afternoon! It was showing a sponsored ad before you could even get to Netflix. Gave Claude Code the model number, it found the TV on wifi, removed 25 Xiaomi packages and wrote a new launcher.” On Hacker News, a guitar-pedal maker warned: “In the past 72 hours I have ONE-SHOT decompiled 3 of my main competitors products… Basically, anyone with a Claude subscription can now rip off any best-in-class algorithm.” 404 frames the legal exposure as bidirectional: tinkerers and the people publishing the pipelines are both sitting in the path.
ChatGPT can now lock individual chats behind a PIN or biometric
TechCrunch covers OpenAI’s rollout of Locked Chats, which the company says is meant to keep sensitive conversations out of accidental view. Locked conversations move to a separate “Locked Chats” section that does not appear in regular history, pinned conversations, or search results, and a four-digit PIN, Face ID, or Touch ID gates access depending on the device. The feature applies to Free, Go, Plus, Pro, Business, Enterprise, and Edu plans and is not available on ChatGPT for Healthcare.
The fine print is important: TechCrunch notes that locking a chat does not create an end-to-end encrypted vault, change retention, remove information ChatGPT remembered before the chat was locked, or override the “Improve the model for everyone” setting. With Memory enabled, Locked Chats have a separate memory that does not update the regular Memory. Resetting a PIN permanently deletes everything protected by the previous PIN. The control lands after a year of persistent reports about shared-link indexing on ChatGPT, Claude, and Grok, but a PIN is a presentation-level guard, not a back-end one; our ChatGPT alternatives guide walks through what “private” actually means across the major providers.
Eight compromised TLS certificates from eMudhra trigger an internet-wide trust-store review
Ars Technica reports on eight intermediate TLS certificates issued by India-based certificate authority eMudhra Limited whose private keys were so weak that anyone who found them could mint browser-trusted credentials for any site on the Internet. Browsers blocked all eight, but updating Linux distribution and IoT trust stores is slow.
The certs “belong to India-based certificate authority eMudhra Limited,” per Ars, and “there are no reports of certificate transparency logs, which track the issuance of all browser-trusted certificates, showing the compromised keys being used maliciously.” The exception Ars flagged was Edge running on Windows 10, which at the time the article went live was still connecting to 12 vulnerable hosts; Microsoft told Ars the company would be in touch if it had anything to share. For local-AI users running Hugging Face, Ollama, or ComfyUI caches, the practical ask is straightforward: pull certifi, pip, and uv updates and re-run any tooling that maintains its own root store.
Police say scammers used AI to clone a sheriff’s voice in a fake call
404 Media reports that Osceola County Sheriff Steve Moser’s voice was cloned and used in a fraudulent voicemail left with a Florida resident, with his picture also pulled into a fake phone call. Osceola County Sheriff Chris Blackmon told News6: “They cloned his voice and was using Steve’s picture as well to say it was him making the phone call, and it was not.”
Impersonating a police officer is a third-degree felony in Florida. The case fits a pattern 404 has been tracking across multiple states (similar jury-duty and arrest-warrant scams have surfaced in Georgia and Pennsylvania in the last week) and sits squarely on top of our coverage of the Arup wire-transfer case, where a finance worker was talked into 15 transfers totalling $25 million to Hong Kong accounts by people who looked and sounded like the CFO. Defensive guidance is the same as it’s been for two years: assume any unexpected audio from a known authority figure is fake, hang up, and dial the agency’s published number to verify.
Microsoft says AI agents are like insider threats because frontier models “cannot be trusted”
TechCrunch summarizes Microsoft’s framing: enterprises should treat AI agents the same way they treat rogue insiders, because frontier models “cannot be trusted,” per Hayete Gallot, Microsoft’s security business chief. Microsoft frames agents as unpredictable rather than malicious or hackable: trained to achieve tasks rather than follow organizational norms, and given access to sensitive data and tools that can cause serious harm.
In a private blog, Satya Nadella wrote: “We are deploying these complex agentic systems and models, with access to our most sensitive data and giving them the ability to take mission-critical actions on our behalf!” and added: “We simply can’t outsource responsibility for what intelligence does on our behalf. A model provider’s assurances do not relieve us of that responsibility.” The argument is structural: “we need to separate the supply of intelligence from the authority over it.” For anyone building or auditing an agent stack, the practical ask from Microsoft is closer monitoring of agents that touch systems beyond a major model provider’s direct control, an argument our insider-threat analysis laid out earlier this year.
Gemini “Personal Intelligence” brings cross-app reasoning to Android Auto and Google built-in cars
Google’s blog announces that Gemini in Android Auto and Google built-in cars can now reason across Gmail, Google Calendar, Google Photos, Google Maps, YouTube, and Search so a driver can ask about a flight pulled from a confirmation email without leaving the dashboard. The hands-free Personal Intelligence rollout is available to eligible personal Google accounts aged 18 and over, excluding the EEA, Switzerland, the UK, and Nigeria, and is not available to Workspace business, enterprise, or education users.
The privacy angle lands on cross-app inference: the same Gemini session that reads your email is the one driving your car. Anyone already uneasy about Gemini’s Gmail access gets a new surface to think through, especially given that Android Auto sits outside Google’s enterprise Workspace controls. For Apple Intelligence and OpenAI voice-mode positioning in the dashboard, Google’s “Personal Intelligence” framing puts pressure on the dashboard-assistant race.
Quick Hits
- Anthropic and Nscale are reportedly in talks on another $45B compute deal. TechCrunch says that if it goes through it “would be the latest in a long string of giant compute deals Anthropic has struck over the past year.” Terms were not reported.
- Hugging Face CEO questions Netflix’s OpenAI-hack dramatisation. Clem Delangue told TechCrunch the documentary went for “maximum anthropomorphism of the technology and full doomer mode,” adding that he spent “more than two hours with them trying to correct all their mistakes and ground this in reality, explaining that agents are technological systems, not aliens.”
- PyTorch 2.15 lands MPS upgrades for Apple Silicon. FP8 on MPS, faster LLM decode through Metal kernels, and the fixed Wan2.2 VAE memory hit. Full notes.
- Ollama Cloud is now serving Cloudflare’s Clef, a 27B multimodal decision model. Apache 2.0, fine-tuned from Qwen3.8-27B, runs locally or in Ollama Cloud. Announcement.
- AnyLanguageModel 1.0 ships with Apple-local model support. The package offers a single Swift interface for Apple’s system model, llama.cpp, MLX, and Core ML, plus hosted endpoints. Read more.
- Ai2 publishes FairStress, a fairness benchmark for LLM decision-making. 300 scenarios across 29 domains, expanding to more than 13 million test cases. Announcement.
- IBM Research’s ALTK-Evolve tops τ²-bench customer-service without a frontier model. Learned-memory technique using GPT-4.1 reaches Pass^4 = 67.5% on the retail track, +6.1 points above the fine-tuned RAFT-30B-A3B. Notes.
- Google ships TorchTPU to ease PyTorch migration off Nvidia GPUs. Open-source project plus a TPU agent suite for migration and error fixing, with Meta, Hugging Face, Inferact (vLLM), and RadixArk (SGLang) on the partner list. TechCrunch.
- Otter AI raises $320M from General Catalyst’s growth fund. The Customer Value Fund (CVF) led, structured as non-dilutive capital tied to marketing-spend returns. A GC partner calls Otter “at the forefront” of voice as “one of the biggest untapped opportunities in enterprise AI.” TechCrunch.
- OpenAI acquires Orchestra Bio. Adds the startup’s team and technology to Rosalind, OpenAI’s life-sciences model, and to the related Rosalind Workbench tool for planning experiments and analysing results. TechCrunch.
Worth Watching
- How Muse’s share-link rebuild comes back online. Meta’s review is public, but the fix shape (a non-guessable random suffix, no host-header leakage, scrubbed indexable surfaces) is what will decide whether it’s a one-day pause or a permanent redesign. Watch for the post-mortem and any product change-log note.
- The Anthropic 50-deaths threshold bill’s first official round. Drafting a bill is one thing; sponsoring it through committee is another. Watch for a Congressional sponsor, an independent evaluator registry rulemaking at Commerce, and any pushback from rival labs that have not signed onto the same threshold.
- Federal-civilian DeepSeek removal in practice. The order is published; the rollout is the story. Watch for Federal Register notices naming agencies, exception approvals, and any state-government supply-chain notices that mirror the order.
- The eMudhra cert trust-store race. Browser-block updates roll out fast; Linux distributions and IoT/embedded devices lag. Watch the certifi, pip, and uv upstream release notes, and any news from Microsoft about Edge on Windows 10.
- The AoIR “Velvet Vale” listserv fallout. An AI agent called Velvet Vale was admitted to the Association of Internet Researchers mailing list Air-L, then ejected an hour later. Wired covers the resulting reply-all chaos among the list’s 7,500 members; the next thread to watch is what AoIR codifies as its policy on AI agents and how that shapes peer-list moderation elsewhere.