Tarferret

Do you know a browser extension can basically steal everything going on in your browser? Honestly, most of us just click “Allow” when the permission popup shows up. We want the tool to work. But with the right (or wrong) permissions, a malicious extension can do a lot more than it claims. It can read every page you visit in real time. Banking sites, email, documents, whatever’s open. It can grab your session cookies — those little tokens that keep you logged in — so the attacker gets in as you without needing your password. It can log every keystroke you type. It can inject code into pages you’re already on, change what you see, or drop a fake payment field on a checkout page. It can even route your traffic through someone else’s server or quietly record conversations you have with AI chatbots. This isn’t theoretical. In April 2026 researchers found 108 malicious Chrome extensions. Telegram tools, video helpers, productivity add-ons — about 20,000 downloads total, all sending dat...

AI companies are quietly adding watermarks to the stuff their models generate. Anthropic detailed how Claude will do it. New models from August 2, 2026 onward embed an invisible watermark straight into the text. You can’t see it, it doesn’t mess with the quality, and it usually survives copy-paste plus light editing. Images and other files get signed C2PA metadata so you can check the origin. Google’s been at this for a while with SynthID. It watermarks Gemini text by subtly shifting token probabilities and does the same across every frame of Veo videos. Same idea machine-readable signal that doesn’t ruin the output. Both moves are partly driven by the EU AI Act rules that kicked in. The marks aren’t perfect (heavy rewriting can kill them), but at least the big labs are starting to leave fingerprints on what they produce. Feels like this is becoming the new normal. Sources: Claude Help Center - How Claude marks AI-generated content Google DeepMind - Watermarking AI-generate...

xAI dropped Grok Bot in early beta a couple days ago. These aren't the usual chatbots. Each one gets its own computer up in the cloud. They log into the same tools and apps you already use email, CRM, whatever and actually finish the job. You can close your laptop and they keep going. Inside the company they were already running them for sales outbound, processing invoices from Gmail, reproducing bugs, the works. Now it's open to SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium people on desktop and iOS. You just message them like a coworker. Show one how you do a multi-step thing once and it remembers the routine. You can spin up a few at the same time and they even talk to each other. Honestly, the "own computer" part is what makes it feel different. Most agents still feel like they're stuck in a chat window. This one works where the real work lives. Still early days, but kinda interesting to watch how this plays out. Sources: xAI - Introduci...

Google just added a new way to get into your account: a selfie video . Announced this past week, it's meant as a backup option if you're locked out or don't have your usual phone or computer handy. Setup is simple enough. You look at your camera and do a few guided head movements — turn left, turn right, nod, whatever it asks — so it captures your face from multiple angles. That becomes your reference video. Later, if you're locked out, you record another one and it gets matched against the original to confirm it's really you. The part I actually care about as a security person: Google says they're running layered checks specifically to catch deepfakes and fake photo/video impersonation attempts, plus requiring "liveness" — those little head movements — so a static photo won't cut it. They're also pairing it with their existing suspicious sign-in detection. On privacy, Google says the video is encrypted at rest, stored under your contro...

Okay this one's wild. A federal complaint against an alleged Scattered Spider hacker just outed a Windows tracking ID most people (including me, honestly) had never heard of: the GDID , or Global Device Identifier. Here's the gist. When you sign into Windows with a Microsoft Account, a server assigns your PC a permanent ID number. It's stored locally in your registry, read by background services like Phone Link and cloud clipboard, and stamped onto activity your PC reports back to Microsoft, including update sharing data. It survives Windows updates. Reinstalling gets you a new one, but Microsoft's own filing admits a single account can rack up multiple GDIDs over time, easily linked back together. In this case, the FBI used it to trace a 19-year-old hacker's device across VPNs, proxy servers, and three countries, tying it to logins on his Snapchat, Apple, and gaming accounts over eight months. He got caught, no complaints there. But the case exposed something ...

Okay this one's interesting. Chinese AI company StepFun just unveiled STEPX Neo, which they're calling the first "large-model-native" AI phone. Not just a phone with a chatbot bolted on — the whole OS, called Step AOS, is rebuilt from the ground up so an AI agent (they call it "Amoo") runs the show instead of you tapping around apps. Here's the part that caught my attention as a security person. Step AOS breaks the entire OS down into tiny callable units — comms, apps, files, system — using something they're calling an MCP-standard "atomic capability engine." Basically the agent can reach into almost anything on your phone and act on your behalf. Book a ride, pay a bill, edit a doc, whatever. StepFun says they've thought about the obvious problem here (an AI with root-level access to your whole digital life is a juicy target) and built in a "trusted, visible, controllable, reversible" framework — sandboxed execution, audi...

Texas just dropped a lawsuit on Meta and WhatsApp, and it's getting some attention. The Attorney General's office says the company has been misleading people about how secure their messages really are. According to the suit, WhatsApp markets itself as fully encrypted and private, but Texas claims Meta can still access "virtually all" private communications. That's a pretty big deal if true, especially for an app billions use daily for chats that feel personal. Ken Paxton called it out directly — they promise security but don't fully deliver. Meta pushed back hard, saying the allegations are false and WhatsApp can't access encrypted messages. Classic he-said-she-said at this point. The lawsuit wants to stop Meta from accessing Texans' messages without consent and hit them with penalties. It references news reports and a whistleblower angle too. Honestly, this stuff makes me think twice about default apps. End-to-end encryption sounds great on p...

I'm already locked in on this one. Android has a real problem with vendors and ads pushing excessive RAM usage. There's almost no proper way to stop apps from running in the background, and the old line "free RAM is a waste" just doesn't help when you can't open multiple apps without everything slowing down. On Windows 11 PC, 4GB RAM is still useful and one can open a few apps without hanging or unresponsive UI. But on Android 13 and above, you can basically only use one app at a time comfortably. Some companies try to fix it with lite editions of Android, but those are mostly useful only for web apps designated as lite versions. Old SoCs get recycled into cheap devices, but actual old phones get abandoned fast with poor support. New phones rarely deliver the hardware value they used to. And now Google is adding another layer with Aluminium OS – an Android-based desktop system for laptops and PCs launching in 2026. From the official site, it ...

This morning things got spicy in the Telegram mod scene. The developer of Nekogram basically admitted that the app collects info about which account is linked to which phone number. Someone (Sota) dropped the Extra.java code fragment, and a few hours later the dev himself said in his chat: “If your question is ‘Is it true?’, the answer is yes, numbers were sent to the bot.” He followed up with something like “what kind of explanation do you need? It is exactly what it looks like; it is what it is.” The dev claims no numbers were actually stored or shared with anyone, and that people might find that hard to believe. Honestly, yeah… it is kinda hard to believe without seeing the full picture. There's also mention of the data going to some “breakthrough” bots, but he didn't really address that part. One of those bots even admitted Nekogram is using their TgDB search bot automatically (without their knowledge or partnership), probably to look up usernames. If ...