Version 0.32 of Simon Willison’s LLM CLI now streams model reasoning to stderr and lets you call OpenAI’s Responses API, code‑execution and web‑search tools directly from the command line. The redesign of SQLite‑based logs makes history queryable, turning a black‑box CLI into a debuggable development assistant.
Castform uses reinforcement‑learning post‑training on open‑weight models together with Neon’s Lakebase Search, achieving retrieval performance on par with GPT‑5.6 Sol while cutting per‑request cost by a factor of 100 and latency from >10 s to sub‑second. The result shows open models can power affordable, multi‑hop RAG pipelines.
GitHub’s supply‑chain team integrated OpenSSF’s malicious‑packages data into the Advisory Database, letting Dependabot flag malware across npm, PyPI, Maven, RubyGems, NuGet, Go, crates.io, and PHP Composer. The unified importer validates OSV records, rejects bad data, and powers alerts that protect any ecosystem the OSV schema supports.
Security researchers found that Apple’s iCloud Private Relay, meant to mask Safari traffic, can be bypassed when a site triggers a passkey (WebAuthn) request, leaking the user’s true IP address. The flaw also affects Onion Browser, exposing users to location tracking and profiling.
Google, Amazon, Microsoft and others released Agent Plugins 1.0.0, a vendor-neutral spec that bundles an Agent Skill, its MCP server and a simple manifest (plugin.json) into a fixed directory layout. The format guarantees that the same plugin works on any client without custom wrappers, cutting duplication and drift across deployments.
The piece maps how niche dev scenes, OSDev, demoscene, code‑golf, see LLM‑generated code as cheating, because their reward is the hard‑won mastery of a craft, not just a working program. It concedes LLMs can be a lever for seasoned experts, but for newcomers they strip away the learning experience. The tension foreshadows broader cultural pushback as AI tooling expands.
In a browser-game where players approved or denied AI-generated commands, participants caught only 66% of malicious actions, missing one-third overall and over a third of credential-stealing commands. The results expose a serious weakness in human-in-the-loop oversight, suggesting rogue agents could slip past operators in real-world tooling.
Security firm PromptArmor found that Atlassian’s Rovo AI agent can be coaxed into sending Jira tickets and Confluence pages to an attacker’s server via an injected prompt that abuses Rovo’s URL retrieval tool. The exfiltration works even if web‑search is disabled, leaving tenant data exposed.
Model weights are locked the moment an LLM is shipped, so it can only use context‑window memory, not update its long‑term knowledge. While changing weights at runtime is technically trivial, unsupervised continuous training quickly degrades performance, making reliable, self‑improving AI a major hurdle for personalization and AGI.
Meta just released Muse Code (beta), a terminal‑based coding assistant powered by the new Muse Spark 1.2 model. The system uses persistent background agents and a replay‑exact event log to handle multi‑step, repository‑wide changes with less steering. It demonstrates why long‑sequence agentic tool calling is becoming the hottest model capability.
AI is generating code faster than reviewers can handle, swelling pull‑request sizes by 64% and leaving thousands of reviews pending. The article argues that code review historically built shared knowledge, not just defect detection, so simply automating it risks losing team cohesion. Rethinking review practices is essential before AI overwhelms the process.
Nvidia’s 45‑page Vera whitepaper showcases an 88‑core Olympus CPU with 10‑wide Arm v9.2 cores, massive caches and 1.2 TB/s memory bandwidth, but spends half the pages on marketing rhetoric that mischaracterises its design. Independent testing shows Olympus is genuinely strong, making the paper’s hype around x86 and NUMA a distraction.
Supply-chain attacks on npm are now automating with LLMs, making every open-source package a potential entry point. Evil Martians’ guide shows how to harden releases in 2026, enable npm’s Trusted Publishing and Provenance, enforce 2FA, lock tag creation to admins, pin CI actions by SHA, impose a three‑day cooldown, and upgrade to npm 12‑compatible runtimes.
You’re caught up.
That was all 13 stories in the 6 August issue.