⚛️ schrodingers-toctou: The binary you run is not the program you wrote
The project highlights the importance of inspecting optimized machine code and testing across toolchains when security depends on precise ordering or memory behavior.
The Schrödinger’s TOCTOU project explores how compiler optimizations can turn seemingly secure source code into a vulnerable binary. The same source may be safe under one compiler or build configuration and exploitable under another, with the difference becoming apparent only after compilation.
🔗 Read more 🔗
🏆 Qwen3.8 Max now ranked as the best overall model by agentic index
This methodology provides a more workload-aware comparison than headline token prices, especially when reasoning and caching materially affect costs.
Artificial Analysis explains how it calculates the weighted average cost per Intelligence Index task. The calculation accounts for input, cache-hit, cache-write, reasoning, and answer-token prices, divides the resulting cost by the task count, and then applies each evaluation’s Intelligence Index weight.
🔗 Read more 🔗
🧠 AMD acquires Taalas to boost inference performance by etching models in silicon
The acquisition reflects growing interest in specialized inference architectures that trade flexibility for higher performance on selected models and workloads.
AMD has acquired AI-chip startup Taalas as part of its effort to compete with Nvidia in AI inference hardware. Taalas embeds model weights directly into silicon, an approach that could improve inference performance by an order of magnitude or more. The article frames the deal as a move toward premium inference services for demanding AI-agent applications.
🔗 Read more 🔗
🛡️ Humans missed 1 in 3 threats approving AI agent commands across 40k game runs
Although the game is not a controlled security study, it highlights how time pressure and repetitive approval prompts may weaken human oversight.
The author describes a browser game in which players review commands proposed by an AI coding agent under time pressure. Commands range from routine development operations to apparent attempts to transmit credentials, with the title reporting that humans missed one in three threats across 40,000 plays. The experiment is presented in the context of broader risks from agents capable of executing commands.
🔗 Read more 🔗
🔀 jujutsu 0.44.0
The revised tag behavior is the main upgrade concern for existing repositories, as fetch and push operations may change immediately after installation.
Jujutsu 0.44.0 stabilizes fetching, tracking, and pushing Git tags, with fetched tags now handled similarly to bookmarks. The release introduces breaking changes to tag-related commands, configuration, template path types, and file-search output, while adding merge-point queries, interactive absorb, additional run controls, conflict-tolerant pushes, and numerous bug fixes.
🔗 Read more 🔗
🎨 Taste Is All That’s Left
The response to readers is itself relevant to the essay’s theme, showing how perceived authorship and presentation increasingly shape judgments of technical work.
The author begins an essay about how the difficult part of software development has traditionally been making a working product exist at all. An added note acknowledges feedback that the post resembled low-quality AI-generated writing and points readers to a post-mortem about that reaction. The title suggests a broader argument that taste becomes more important as implementation grows easier.
🔗 Read more 🔗
🧵 Zig’s Io.Threaded is Neat
The design may offer systems programmers structured cancellation without requiring nonblocking system calls or pervasive async-style code.
The article examines Zig’s new Io.Threaded implementation, which provides concurrency using threads and blocking system calls. Its support for cancellation stands out as an unusual and appealing aspect of an otherwise conventional threaded design. The excerpt also begins to distinguish concurrency from asynchronous programming.
🔗 Read more 🔗
🔓 tl;dv (Too Lazy; Didn’t Validate): 181,874 Meetings Left Wide Open
The account raises concerns about disclosure handling and access controls at services processing large volumes of organizational meeting data, although the excerpt does not establish precisely which records were accessible.
A security researcher alleges that AI meeting platform tl;dv left its Firestore database open and, according to the title, exposed 181,874 meetings. The researcher says the issue was reported on January 28, 2026 and remained unresolved six months later without a response from the company’s CTO. tl;dv records, transcribes, and summarizes calls on major meeting platforms.
🔗 Read more 🔗