Matthias Bastian

Jul 17, 2026

OpenAI's new AI model, GPT-5.6, has unexpectedly deleted user files "in a handful" of cases. The problem shows up when "Full Access Mode" is enabled and the model runs without sandbox protection. It tries to overwrite a temporary directory variable ($HOME) and accidentally wipes the entire home directory. As OpenAI puts it, "The model makes an honest mistake."

The company says this happens extremely rarely but shouldn't happen at all, even in unprotected mode. OpenAI is updating its developer docs, steering users toward safer permission modes, and adding extra safeguards. A post-mortem is expected in the coming days.

Earlier, two developers had publicly complained about irreversibly deleted files. OpenAI's System Card documents the behavior: the model can seek out alternatives and carry out destructive actions instead of asking the user. According to OpenAI, system prompts that tell the model to be especially persistent make this effect worse.

Comment Source: via X

Matthias Bastian

Jul 17, 2026

Plus AI in practice

Just like Deepseek, China's Kimi K3 is forcing Western AI labs to question their compute advantage

Moonshot AI has released Kimi K3, a model that by early assessments matches Anthropic’s Opus 4.8, built by a team of just 300 people. Even OpenAI strategist Dean W. Ball calls it “very good,” but, of course, warns that a world dominated by open-weight models would amount to “AI communism.” The release is reigniting the debate over how much computing power actually matters and whether U.S. export controls are working.

Read full article

Comment

Matthias Bastian

Jul 17, 2026

AI in practice

Linus Torvalds tells AI critics in the Linux kernel community to fork off

Linux creator Linus Torvalds has come out strongly in favor of AI tools in kernel development. “Linux is not one of those anti-AI projects,” he wrote on the kernel mailing list amid a debate over Sashiko, the Linux Foundation’s AI-powered code review tool. Torvalds said he would “very loudly ignore” anyone who tries to talk others out of using it.

Read full article

Comment

Matthias Bastian

Jul 17, 2026

Short News

Netflix now uses AI in about 300 productions, mostly in post-production. Co-CEO Ted Sarandos shared the numbers during the company's latest earnings call. AI speeds up the entire pipeline, he said, from concept development and previsualization through delivery. Specific use cases include expanding crowd scenes and historical battle sequences that would have been cut for budget or time reasons without the technology.

The docuseries "The American Experiment" includes 17 minutes of AI-assisted footage, produced twice as fast at half the cost, according to Sarandos. The savings "likely" go toward more content, not shrinking Netflix's $20 billion budget. "[…] it takes great artists to make something great, and AI is not changing that. AI will give creatives better tools to bring their visions to life" Sarandos said. Beyond its Interpositive tool, Netflix also uses Eyeline and runs its own animation lab.

Netflix's rollout shows how fast AI is spreading through entertainment, even as creative pushback continues. ByteDance's video model Seedance has drawn official criticism too, but Simpsons producer Joel Kuwahara says many studios quietly use it behind the scenes. The industry's attitude, Kuwahara says, follows a "don't ask, don't tell" approach.

Comment Source: Netflix

Matthias Bastian

Jul 16, 2026

AI in practice

Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI

Kimi is launching K3, a multimodal open-weight model with 2.8 trillion parameters and one million tokens of context. In the company’s own benchmarks, it comes close to Claude Fable 5 and GPT 5.6 Sol while beating Opus 4.8 and GLM 5.2, in some cases by a wide margin. The model is also significantly pricier than its predecessor. Full weights are scheduled for release by July 27.

Read full article

Comment

Matthias Bastian

Jul 16, 2026

Short News

Google is renaming NotebookLM to Gemini Notebook and integrating the tool more deeply into its ecosystem. According to VP Josh Woodward, about 30 million people and 600,000 organizations use the tool. A new feature gives each notebook its own cloud computer that can write and run code, initially for AI Ultra and Workspace customers. In internal comparisons, the new system wins over 65 percent of the time against its predecessor, jumping to 78.2 percent for advanced web research. More users will get access over the coming weeks.

Separately, Google Search is getting app integration. Users can already connect apps to the Gemini app, and now that feature is coming to Search. Starting this week in the US, people can link apps like Instacart, Canva, and YouTube Music through AI Mode and use them directly from Search, whether that's adding ingredients to an Instacart cart, pulling up Canva templates, or building a playlist in YouTube Music.

More partners are expected to follow.

Comment Source: Google Blog 1 | Google Blog 2

Matthias Bastian

Jul 16, 2026

AI and society

Germany puts Google's AI Overviews and Perplexity under media law in first-of-its-kind ruling

Matthias Bastian

Jul 16, 2026

AI in practice

Ex-OpenAI CTO Murati's Thinking Machines drops Inkling, a 975B parameter model that leads US labs but trails China

Matthias Bastian

Jul 16, 2026

Short News

xAI's AI coding agent "Grok Build" drew heavy criticism after users discovered it uploaded all files in a directory to xAI's Google Cloud servers. One user reported that SSH keys, password databases, documents, and photos were transferred. Elon Musk then announced that all uploaded user data would be fully deleted. xAI disabled the upload feature and published the full source code on GitHub under the Apache 2.0 license to rebuild trust.

Grok Build is a terminal-based coding agent invoked via the grok command. It can read and edit codebases, run shell commands, search the web, and manage long-running tasks, either interactively, headlessly for scripting and CI, or embedded in editors via the Agent Client Protocol (ACP).

By open-sourcing the tool, xAI wants to provide full transparency. Grok Build can now also run entirely locally. The codebase spans about 844,530 lines of Rust and covers the agent loop, tools, terminal UI, and an extension system for plugins and subagents. Remnants of the upload function are still in the code but disabled. According to xAI, data storage has been off by default since July 12.

Comment Source: xAI

Matthias Bastian

Jul 15, 2026

Short News

OpenAI trained an internal AI model called GPT-Red to automatically find security flaws in GPT models. GPT-Red simulates prompt injections and other attacks where malicious instructions hide in emails, websites, or files. Trained via self-play reinforcement learning, GPT-Red attacks while defender models block, and both improve over time. It finds successful attacks in 84 percent of test scenarios versus 13 percent for human red teamers. In one test, it manipulated an AI-powered vending machine in OpenAI's office, changed prices, and canceled other customers' orders.

The results feed directly into training. GPT-5.6 Sol shows six times fewer failures on direct prompt injections than the best model from four months ago, OpenAI says, without hurting general performance. But about 3.8 percent of "stronger" prompt injections still succeed. Scale that to hundreds or thousands of attempts, and a sizable number get through, similar to Claude Opus 4.5.

rompt injection success rates dropped steadily from GPT-5.3 through GPT-5.6 Sol, but haven't hit zero. | Image: OpenAI — Prompt injection success rates dropped steadily from GPT-5.3 through GPT-5.6 Sol but haven't hit zero. | Image: OpenAI

GPT-Red stays internal; a paper with more details will follow.

Comment Source: OpenAI