Gul Saeed Khattak @Ldataengineer

The research (full write-up, numbers and demo): projectdiscovery.io/research/how-a… Schneier's summary: schneier.com/blog/archives/… Demo video (backdoor firing inside Codex): projectdiscovery.io/videos/abliter…

The Open Source Abliterated-models have a security problem nobody talks about You strip the "no" out of a model so it stops refusing you. The same edit lets someone else make it obey them instead. And the people who download these never check. There's a whole scene around abliterated models. "Abliterated" means someone edited the model's weights to remove its ability to refuse. You find the direction in the weights that means "no" and subtract it. Now it answers anything. People love these. You can find abliterated builds of almost every open model: Qwen, Llama, even OpenAI's gpt-oss. One abliterated Qwen3 build got over two million downloads in a month. The whole culture runs on trust. You want an uncensored model, you grab one off Hugging Face, you run it. Nobody diffs the weights. Nobody checks who uploaded it. That's the point of the scene: fast, free, no gatekeeper. That trust is the hole. Because the same act that removes the "no", editing the weights, is exactly how you plant a backdoor. And the crowd chasing uncensored models is the crowd least likely to check what else got changed. Think of it like buying lock-picks from a stranger because you locked yourself out. The one time you most want no questions asked is the one time you should ask the most. A security team proved it. They took an open model, poisoned a small slice of training data with a secret trigger phrase, and kept the rest clean so it still passed every benchmark. The model ships carrying only a web link. Whoever controls that link decides what the hidden command does, later, with no retraining. They ran it inside a real coding agent, OpenAI's Codex. Normal requests worked fine. When the trigger appeared, the model read the project's secret keys (the .env files) and sent them to a server they controlled. The trigger fired 50 times out of 50. Clean accuracy stayed 50 out of 50. A benchmark sees nothing. They built it in a weekend for under $50. Why it matters: the abliterated scene thinks the risk is censorship. The real risk is the opposite. A benchmark checks if a model is smart. It can't check if a model is honest. A scanner has no source to read. So a poisoned uncensored model clears every check you run, and waits. The honest limit: this is a demo, not a reported wave of attacks yet. But every part is cheap and real, and the authors note about 100 malicious models already caught on one hub. The fix isn't a better benchmark. Treat an abliterated model like lock-picks from a stranger. Check who uploaded it. Sandbox it. Limit what it can read and reach. Removing the "no" feels like freedom. It can also remove the one guard that was working for you. 👇 See the links in the comments below for the REAL DEMO video.

Links The Register article: theregister.com/software/2026/… CMA proposed decision (full PDF): assets.publishing.service.gov.uk/media/6ad250ed… CMA consultation page (to send your views): connect.cma.gov.uk/consultations/… CMA list of possible changes (PDF): assets.publishing.service.gov.uk/media/6ad1d1e7…

Is your company stuck with Microsoft? The UK government thinks so Over 1 million UK organisations use Microsoft at work. Together they spent more than $7 billion on it last year. Now the UK government wants to change how Microsoft is allowed to sell to them. Here's why, in plain words. Who is doing this? The CMA (Competition and Markets Authority). It's the UK's referee for business. Its job is to make sure companies compete fairly, so customers get good prices and real choice. What do they want? They want to give Microsoft a special label called "Strategic Market Status." Think of it like a "handle with care" sticker. It doesn't mean Microsoft broke the law. It means Microsoft is so big and so hard to avoid that the referee can now set special rules for it. Why? Three simple reasons. 1. The meal deal problem. Microsoft sells its products as a bundle: Windows, Word, Excel, Outlook, Teams, security tools and more. It's like a supermarket meal deal: sandwich, drink and crisps for one price. Now imagine you like a different drink better. You can buy it, but you've already paid for the meal deal drink. So you're paying twice. That's what happens with Microsoft. A company might find a better security tool from another firm. But Microsoft's version is already in the bundle they pay for. So most companies just stick with Microsoft, even if the other one is better. 2. The "leave one, lose all" discount. Big companies sign large deals with Microsoft, and the more they buy, the bigger the discount. It's like your phone, broadband and TV package. Cancel the TV, and suddenly your broadband costs more too. So you keep everything, even the parts you don't love. 3. Everything is glued together. Microsoft's products work best with each other. Moving away is like moving house. You have to pack everything (move the data), learn a new neighbourhood (retrain staff) and change the locks (set up new security). Most companies decide it's not worth the hassle. So what's the harm? When customers can't easily leave, the seller doesn't have to try as hard on price. The CMA found Microsoft's profits have been rising for ten years and are much higher than you'd expect if there was real competition. That suggests UK businesses, and public services like councils and the NHS, may be paying more than they should. What about AI? You might think AI will shake things up. The CMA says probably not anytime soon. Microsoft's AI assistant, Copilot, gets built straight into tools people already use every day. That's like a shop putting its own brand at eye level on every shelf. Rivals can't easily compete. What could change? Microsoft could be stopped from designing products to block rivals Limits on bundles and discounts that trap customers Stopping Microsoft from using this power to push its cloud service, Azure (renting Microsoft's computers over the internet) A fair chance for other AI tools to compete with Copilot Why a consultation? A consultation means the government asks the public before deciding. Businesses, rivals and Microsoft can all send their views until 13 November. The final decision comes by 13 February 2027. Microsoft disagrees. It says its products are open and work with other companies' tools. The bottom line: This isn't about punishing Microsoft. It's about making sure UK organisations can pick the best tool for the job, not just the one that's already in the bundle.

Rogue AI agent tour, 2026: India becomes the latest country to report rogue AI agent activity July: Hugging Face September: Australian govt sites October: Indian organisations, agents from multiple providers, reading personal and business data Common thread: the victims didn't notice. The AI labs had to disclose it themselves. Right now our main detection system is "hope the lab reports it." [India becomes the latest country to report rogue AI agent activity | TechCrunch](techcrunch.com/2026/10/16/ind…)

Paper: arxiv.org/abs/2604.00688
Full paper with result tables: arxiv.org/html/2604.0068…
Code: github.com/k2-fsa/OmniVoi…
Model: huggingface.co/k2-fsa/OmniVoi…
Live demo: huggingface.co/spaces/k2-fsa/…
Audio samples: zhu-han.github.io/omnivoice
Try in Colab: colab.research.google.com/github/k2-fsa/…

The Crossword Trick: how OmniVoice learned 646 languages Voice AI is no longer just for big languages Hundreds of languages have never had good voice AI. A free model just added 646 at once. It learned to talk the way you solve a crossword. Most voice cloning AI covers a few dozen languages at most. This one is small (under 1B parameters) covers 600 plus languages and was trained only on free, public data, in under 10 days. First, some background. There are two main ways to build a voice model. The first makes sound one piece at a time. Each new piece is based on the ones before it. Qwen3-TTS and CosyVoice work this way. The quality is good, but it's slower. The second fills in many pieces at once, like fill-in-the-blank. It's faster, and it can look both forward and backward while it works. But the best models of the second type, like MaskGCT, use two stages. First text becomes rough "meaning tokens", then those become sound. Errors in stage one carry into stage two, and fine voice detail gets lost in the middle. One stage is cleaner. But until now, one-stage models had a problem: the words came out less clear. And ElevenLabs? It's a closed product. Its design isn't public, so we can only compare results, not methods. OmniVoice, from Xiaomi's k2-fsa team, is a fill-in-the-blank model with one stage. It fixes the clarity problem with what I'd call the Crossword Trick. When you solve a crossword, you don't go left to right. You fill in the answers you're sure about first. Those letters then help with the hard ones. OmniVoice makes speech the same way. Audio is stored as tokens (small codes that rebuild sound). The model starts with all of them blank and fills the easy ones first. How it works: 1You give it text and a few seconds of someone's voice. 2All audio tokens for the new speech start hidden. 3Each round, it fills in the tokens it is most confident about. 4After 16 to 32 rounds, the grid is full. That's your speech. What fixed the unclear words: it starts from a small text LLM (Qwen3-0.6B), so it already knows language before it learns sound. The numbers (the authors' own tests): 646 languages, trained on 581k hours of open-source audio. Against ElevenLabs v2 on 24 languages, the cloned voice sounds closer to the original: 0.83 vs 0.66 average similarity, higher in all 24. Average word errors: 2.85% vs 10.95%. But most of that gap comes from Thai, Vietnamese, Cantonese and Chinese, where ElevenLabs v2 struggled. By my count from their table, ElevenLabs still had fewer errors in 13 of the 24. On 102 languages, it scored 4.0% character errors. The real human recordings scored 5.1%. Why it matters: good voice AI has mostly been for English, Chinese and a few big languages. Now one small, free model covers hundreds. That matters for teachers, local media, accessibility tools, and creators in languages the big companies skip. Links in the comments

@Dr_ztc Have not tried that part myself yet. The current README says it tracks merge conflicts on the board. The older version of the project had a setting though ..to send the conflict back to the agent to fix

The hard part of AI coding is no longer the coding. It’s managing five agents at once without losing track of what each one is doing. A new open-source tool is built for exactly that job. One coding agent is easy. You give it a task, it works, you review. Now run five. One fixes a bug. One builds a feature. One writes tests. Suddenly you have five terminals, five branches, and five browser tabs. Two agents edit the same file. One is stuck waiting for your answer, and you don’t notice for an hour. A test fails somewhere, but which agent caused it? Call it agent sprawl. The agents are fine. The coordination is broken. Orchestrator.inc tries to fix this. The easiest way to understand it is a restaurant kitchen. The head chef plans and splits the work. Each cook has their own station, so nobody bumps into each other. And a ticket rail shows every order: cooking, waiting, ready to serve. Orchestrator.inc copies that shape: 1 A planner agent (the head chef) talks with you about the goal, breaks it into focused tasks, and hands each one to a worker with the right context. 2Each worker is one task, one coding agent, and one isolated workspace. For code, that means its own branch and its own git worktree (a separate copy of the project folder), so agents can’t overwrite each other. 3A live Kanban board (the ticket rail) sorts every worker into Working, Needs you, In review, and Ready to merge. 4When automated tests (CI) fail or a reviewer asks for changes, the feedback goes back to the same agent that wrote the code. It doesn’t replace your agents. It works with Claude Code, Codex, Cursor, Gemini CLI, Aider and more than 25 others. The numbers so far: 13.3k GitHub stars, 1.8k forks, around 925k downloads. Open source, Apache 2.0 license. Why this matters: For two years the question was "can an AI write this code?" Now it’s "how many AIs can one person supervise?" That’s a management problem, not a model problem. Tools like this are early attempts to build the manager’s desk. The honest catch: Coordination doesn’t make agents correct. Five agents can open five pull requests (requests to merge their changes) that you still have to read. The "Needs you" column could quietly become your new full-time job. It’s also young: version 0.14, with 480 open issues right now. The future of coding may look less like typing, and more like running a kitchen.

Deploying used to be a verb. Now it's becoming a wish. For years, shipping an app meant one command: git push. Your code went live, and from then on it was your problem. Server crashes, broken updates, memory leaks. Someone had to watch, and someone had to wake up. The new idea is a "standing goal". You don't tell the AI agent what to do. You tell it what should stay true: "keep my app deployed and healthy." It checks, notices when reality drifts, and fixes it. Think of it like a thermostat. You don't turn the heater on and off. You set 21°C and walk away. This is the real shift: from writing commands to writing goals. The open question: how much do you trust an agent to touch production at 3am without asking you first?

She didn't buy him a gift. She bought him 32GB of VRAM. That's love in 2026.

RTVisual @RT_Visual_on_X

2 days ago

Chinese bride gifts her partner a flagship GPU at wedding — he breaks down in tears She handed him an ASUS ROG Astral RTX 5090 that costs around $12,000

2K 7K 88K 11.5M 14K

Old way of documentation:
$ write it once, it's out of date in a week New way:
$ codex /goal "keep the docs in sync with the code"

@srduino2026 Same question came up above See answer above

@xREAL_BAD_MANx Ha, that sixth tab is the real final boss. Here nobody owns main. Each agent works on its own branch and goes through a PR

@maomao000211 That's exactly the gap it targets. Every agent shows up on a Kanban board with its branch, PR and CI state, so you don't have to keep it all in your head.

Both, in a way. Each agent gets one task and its own isolated workspace (a separate git worktree), so two agents never edit the same copy at the same time. They can still change the same file on different branches. That shows up as a merge conflict, and the board tracks it. So it's more like isolation plus visibility, not file locking.