litany @US490

@basithladoo Cancer

@k1rallik Was able to disable the flag. Didn't have any app data but did find in program files optimizer model dll, so deleted and uninstalled chrome. Been on brave and Firefox for years anyways

🚨Chrome silently installed a 4GB AI model on your computer. Delete it and Chrome puts it back. Privacy researcher Alexander Hanff verified it. No consent prompt. No notification. The file re-downloads if you delete it. Tom's Hardware and TechSpot confirmed it. Hanff set up a fresh Chrome profile on a clean machine. Never clicked anything. Never typed a single keystroke. Just opened the browser and walked away. 14 minutes and 28 seconds later, Chrome had scanned his hardware, evaluated his GPU, RAM, and storage, then quietly wrote a 4GB file to his hard drive. No permission dialog. No checkbox. No notification. The file is called weights.bin. It contains the weights for Gemini Nano, Google's on-device AI model. It lives buried inside a folder called OptGuideOnDeviceModel inside your Chrome profile. Hanff verified the entire install chain using macOS filesystem logs that Chrome cannot modify. The visible AI feature is a lie. Chrome 147 added a prominent "AI Mode" pill to the address bar. Users assume their queries stay on the device since the 4GB model is sitting right there on the disk. They do not. The AI Mode pill routes every query to Google's cloud servers. The on-device model powers buried features most users will never find. You pay the storage. You pay the bandwidth. The headline feature ignores the local file completely. To check if Google did this to your machine. On Windows: C:\Users\[YourName]\AppData\Local\Google\Chrome\User Data\Default\OptGuideOnDeviceModel. On Mac: ~/Library/Application Support/Google/Chrome/Default/OptGuideOnDeviceModel. If there is a file called weights.bin, the answer is yes. Stopping it permanently requires more than a delete. Type chrome://flags. Search "optimization-guide-on-device-model" and disable it. Search "prompt-api-for-gemini-nano" and disable that too. Restart Chrome. Then delete the folder. Skip the flag step and Chrome re-downloads the 4GB file on the next launch. Chrome has 2 billion users. Hanff estimates pushing 4GB to a fraction of those users represents exabytes of data transfer and between 6,000 and 60,000 metric tons of CO2 emissions per model push. The environmental cost of one company deciding what gets installed on two billion machines. This violates the EU ePrivacy Directive Article 5(3). The same law that requires cookie consent banners. It mandates "prior, freely-given, specific, informed, and unambiguous consent" before storing data on user devices. Chrome failed every part of it. Firefox requires opt-in for AI. Apple Intelligence requires consent. Chrome just took your hard drive. Google did not ask. They took your storage. Your bandwidth. Your electricity. They just took it. Source: Alexander Hanff, thatprivacyguy.com/blog/chrome-si… Verified by: Tom's Hardware, TechSpot, Cybernews, gHacks

do you understand what just happened to your computer.. Google Chrome secretly downloaded a 4GB AI model onto your device. Without asking.. Without telling you.. It's called weights.bin. It lives deep in your system folders. It powers Gemini Nano - Google's on-device AI. And if you delete it? Chrome re-downloads it automatically. Like nothing happened. Just Google deciding your hard drive is their storage unit. At 1 billion Chrome users - that's 4 BILLION gigabytes of data pushed silently across the internet. The carbon footprint alone equals tens of thousands of cars running for a year. Check your disk right now: 📁 %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel To stop it: chrome://flags → disable Optimization Guide On Device Model → restart Chrome → delete the folder. Reshare so people know what's sitting on their computers.

Pirat_Nation 🔴 @Pirat_Nation

2 months ago

Google Chrome is quietly downloading a roughly 4 GB AI model to many users’ computers without clear upfront consent. The file, called weights.bin, is part of Google’s Gemini Nano on-device language model and lands in the browser’s user data folder under OptGuideOnDeviceModel.

keyboard_arrow_left Previous keyboard_arrow_right Next

524 5K 18K 9.3M 17K

Google Chrome is quietly downloading a roughly 4 GB AI model to many users’ computers without clear upfront consent. The file, called weights.bin, is part of Google’s Gemini Nano on-device language model and lands in the browser’s user data folder under OptGuideOnDeviceModel. It powers built-in AI tools such as “Help me write,” smarter tab suggestions, on-device scam detection, and page summarization. The download triggers automatically for devices meeting minimum hardware requirements, and Chrome often replaces the files if deleted. While the model processes data locally, installation happens in the background with minimal notification. The scale is noteworthy. Hundreds of millions or billions of installations add up to thousands of tonnes of carbon emissions globally from data transfer, even though each is a one-time event. To prevent or remove it, go to chrome://flags, disable the entries for the optimization guide on-device model and Prompt API, restart the browser, and manually delete the folder.

keyboard_arrow_left Previous keyboard_arrow_right Next

VLAD TARASENKO PULLS OFF THE FORSBERG OH MY WHAT A NASTY GOAL 😵🤢

🧵 Why the obsession with p < 0.05 is hurting science. A meme. A truth. A reality check.

Robert Sapolsky is a Stanford neuroscientist who proved chronic stress is the silent killer doctors ignore. On Chris Williamson's podcast, he revealed 10 "normal" habits you do every day that wreck your sleep, mood, and nervous system: 1) Replay conversations in your head

We're launching the Anthropic STEM Fellows Program. AI will accelerate progress in science and engineering. We're looking for experts across these fields to work alongside our research teams on specific projects over a few months. Learn more and apply: job-boards.greenhouse.io/anthropic/jobs…

Anthropic: 250 Documents Can Permanently Corrupt Any AI Model Someone can permanently corrupt any AI model in the world right now. Not by hacking it. Not by breaking its security. By publishing 250 documents on the internet. That is the finding from Anthropic, the UK AI Security Institute, and the Alan Turing Institute — released in October 2025 as the largest data poisoning study ever conducted. Here is what data poisoning actually means. Every AI model learns from billions of documents scraped from the internet. If someone can plant corrupted documents in that pool before training begins, they can secretly teach the model to behave in specific harmful ways when it encounters a particular trigger phrase. The model learns the backdoor during training. It carries it forever. It does not know it is there. Researchers have known about this attack for years. The assumption was that it required controlling a large percentage of training data — millions of documents — to work on a big model. The bigger the model, the more poisoning you would need. This study proved that assumption completely wrong. The researchers trained models of four different sizes — from 600 million to 13 billion parameters. They slipped in either 100, 250, or 500 malicious documents. Each poisoned document looked like a normal web page at first — a short extract of legitimate text — and then contained a hidden trigger phrase followed by gibberish. 100 documents: insufficient. The backdoor did not reliably form. 250 documents: success. Every model, at every size, was permanently backdoored. 500 documents: same result as 250. The number was constant regardless of model size. A model trained on 260 billion tokens needed the same 250 poisoned documents as a model trained on 12 billion. Scale offered zero protection. Anthropic's own words: "This challenges the existing assumption that larger models require proportionally more poisoned data." Then came the sentence that should end every conversation about AI safety: "Training is easy. Untraining is impossible." Once a backdoor is in the model, it cannot be removed without starting training completely from scratch. You cannot identify which 250 documents caused it. You cannot surgically extract the corrupted behavior. You must rebuild the entire model from the beginning. Anyone can publish content to the internet. Academic papers. Blog posts. Forum discussions. Product descriptions. If even a small fraction of that content is deliberately corrupted before a training run begins, the model that learns from it carries the damage permanently and silently. GPT-5. Claude. Gemini. Every model trained on public internet data is exposed to this attack vector. The defense does not exist yet. The researchers published this not to cause panic — but to force the field to take it seriously before someone uses it. Source: Anthropic, UK AISI, Alan Turing Institute (2025) · anthropic.com/research/small… · aisi.gov.uk/blog/examining…

To check if your Google Workspace has been compromised by the same tool that compromised Vercel: 1. Go to admin.google.com/ac/owl/list?ta… - This is Google Admin Console > Security > Access and Data Control > API Controls > Manage app access > Accessed Apps 2. Filter by ID = …v79i7bbvqj.apps.googleusercontent.com - This is the ID of the compromised OAuth app If you see an app after filtering, you have potentially been compromised

7 everyday things you should never consume: 1/ Protein bars… 🧵

AMD Senior AI Director confirms Claude has been nerfed. She analyzed Claude's session logs from Janurary to March: > median thinking dropped from ~2,200 to ~600 chars > API requests went up 80x from Feb to Mar. less thinking and failed attempts meaning more retries, burning more tokens, and spending more on tokens > reads-per-edit dropped from 6.6x → 2.0x. model stops researching code before touching it. > model tried to bail out or ask "should i continue" 173 times in 17 days (0 times before March 8). > self-contradiction in reasoning ("oh wait, actually...") tripled. > conventions like CLAUDE.md get ignored because there's less thinking budget to cross-check edits > 5pm and 7pm PST are the worst hours, late night is significantly better. this means the thinking allocation is most likely GPU-load-sensitive.

Virtualization vs Containerization 𝗩𝗶𝗿𝘁𝘂𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 (VMs) gives each workload its own full machine with its own guest OS and kernel. Great for isolation and OS flexibility, but you’re paying the cost of booting and managing entire operating systems for every workload. 𝗖𝗼𝗻𝘁𝗮𝗶𝗻𝗲𝗿𝗶𝘇𝗮𝘁𝗶𝗼𝗻 runs workloads as isolated processes on a shared OS. It’s fast and efficient; but with weaker isolation and more constraints around security and OS compatibility. VMs give you strong boundaries and full environments. Containers give you speed and efficient delivery. If you’ve used VMs, you know the pain isn’t the VM; it’s everything around it: provisioning, networking, exposure, and persistence. That’s where exe[.]dev comes in. It gives you instant VMs over SSH, with: • Built-in HTTPS exposure • Persistent environments • Zero cloud config overhead • Safe sandbox execution layer for agents (with secure credential injection via Integrations) When infra is this fast to spin up, you stop hesitating and start testing. Check it out → lucode.co/exe-dot-dev-z7… What else would you add? —— ♻️ Repost to help others learn and grow. 🙏 Thanks to @ssh_exe_dev for sponsoring this post. ➕ Follow me ( Nikki Siapno ) to become good at system design.

Most PhD students think tense errors are a grammar problem. After examining 45+ theses, I can tell you they are a thinking problem. And examiners treat them that way.

Defensive Linux tools

Important takeaways from Claude’s source code: 1. Much of Claude Code’s system prompting is in the source code. This is actually surprising. Prompts are important IP, and I would have thought a sophisticated organization like Anthropic would have performed much or all of their prompt assembly in the server-side harness. 2. Claude Code uses axios, which was also just hacked. Reminder: supply chain attacks are part of closed-source distribution too, and you won’t even know what version of an affected package is being used. 3. The source has a lot of really good comments. These are obviously not for human consumption but for LLMs to understand the purpose of various chunks of code. In the code autocomplete era, most of us engineers hated how many comments were left by LLMs, but perhaps we’ve overcorrected. This looks like a great way to provide context to code outside of the AGENTS.md/CLAUDE.md files. 4. Most folks already know this, but less tools == better results. CC has < 20 tools in normal coding: AgentTool, BashTool, FileReadTool, FileEditTool, FileWriteTool, NotebookEditTool, WebFetchTool, WebSearchTool, TodoWriteTool, TaskStopTool, TaskOutputTool, AskUserQuestionTool, SkillTool, EnterPlanModeTool, ExitPlanModeV2Tool, SendMessageTool, BriefTool, ListMcpResourcesTool, and ReadMcpResourceTool. 5. The “Bash” tool is the crown jewel of Claude Code. A significant amount of deterministic parsing and processing occurs to determine the “type” of commands being run. 6. For better or worse, Claude Code is *all* TypeScript/React with rather explicit Bun bindings. 7. Just because the source is now “available” *DOES NOT MEAN IT IS OPEN SOURCE*. You are violating a license if you copy or redistribute the source code, or use their prompts in your next project! Don’t do that! My overall takeaway: it’s a really well laid-out codebase that is carefully organized to let agents work on it effectively. Direct human intervention here is minimal, but, like with all good projects, the human engineering is still apparent. I’m a bit surprised by some of the shortcuts Claude Code makes, like its prompt assembly being rather messy. Perhaps they have tooling on their side that helps with this introspection, but as it stands, it seems LLMs would struggle to iterate on the prompting because it’s not evident how a given set of parameters assembles a prompt without actually running it. It’s also surprising that the prompts are even in this source code. Keep in mind that even though this is the first time we’ve gotten a proper full-source dump, it has never been impossible to read Claude Code’s prompting since it was part of the actual distributed package — that’s surprising. There might still be a lot of prompting on the server that also gets added (unclear at this point), but there is certainly more than I would have expected in the CLI tool itself.

@lukOlejnik Unfortunately send in the troops. What a disaster, getting played by Netanyahu. Be hard to say look at this for elections with WW3 looming

🚨‼️ We're in contact with the actor behind the Trivy and LiteLLM hack. They told us they are currently extorting several multi-billion-dollar companies from which they've exfiltrated data. They've obtained 300 GB of compressed credentials and are working their way through them as we speak. The LiteLLM compromise alone led to half a million stolen credentials, according to the threat actor. Their message to the world: "TeamPCP is here to stay. Long live the supply chain." They've sent us their new logo (see image) and also teamed up with several threat actors, including Xploiters and Vect.