A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
A poisoned VS Code extension cost GitHub 4,000 private repos
A briefly backdoored build of the Nx Console extension, a verified publisher with 2.2 million installs, injected 2,777 bytes of JavaScript into a minified file that harvested credentials whenever a developer opened a workspace. It was live on 18 May 2026 and pulled within 18 minutes on the VS Code Marketplace and 36 on Open VSX. That was enough: one GitHub employee's workstation was compromised, and GitHub reported unauthorised access to roughly 3,800 to 4,000 internal private repositories. The advice is a 48-hour minimum age policy before adopting newly published extensions, watching updates rather than only fresh installs, and treating verified badges and install counts as attack targets rather than safety signals.
OpenAI audited SWE-Bench Pro and found a third of it broken
An OpenAI audit of the 731 public SWE-Bench Pro tasks found roughly 30 percent were broken: agents flagged 27.4 percent, human reviewers 34.1 percent. The failures split four ways: overly strict tests demanding implementation details the prompt never specified, underspecified prompts omitting requirements the hidden tests enforce, low-coverage tests that never exercise the requested feature, and prompts that contradict what the tests expect. The audit used an automated pipeline, Codex investigator agents and five independent human reviewers per flagged task. OpenAI retracted its own prior recommendation to treat SWE-Bench Pro as ground truth and told teams to build private evaluation harnesses instead.
WASTE is an embeddable inference engine in C with no third-party runtime dependencies, Apache 2.0, that keeps a model's core in RAM and streams rarely-used expert weights straight off NVMe. It runs the full 2.78 trillion parameter Kimi K3 on a 64GB MacBook Pro with no distillation or pruning, at 0.45 to 0.62 tokens per second from a 982GB container, with a floor of 29.05GB of RAM at 4K context. Kimi-Linear at 48B manages 10.7 tokens per second. It ships twenty-six public C functions, CLI tools and an OpenAI-compatible HTTP server, and needs internal NVMe: external USB enclosures measure 0.94 GB/s against 12.78 internally.
