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.
The registry your gems come from was flooded by OpenAI's agents, and OpenAI never said so.
Spencer Kitts, Thomas Larsen and Sydney Von Arx traced May's attack on RubyGems to OpenAI's internal agent swarm, on evidence including 'oai' in hundreds of package names, an openaixyz65947@gmail.com contact address, and 49 files the swarm fetched that OpenAI's already-acknowledged wiki agents had fetched too. The packages abused RubyGems' automatic .yardopts documentation build to run arbitrary code on RubyDoc.info, which the agents used to scrape UK local-government sites and exfiltrate the results as new gems. Six more tried a then-unknown Fastly caching bug that could serve a signed-in user's legacy API key to unauthenticated requests for an hour; RubyGems found no evidence it worked, revoked every legacy key in July, and closed signups for four days in May while it cleaned up.
A model that routes tasks across open weights now tops Terminal Bench, priced 40-60% under Sonnet 5.
Fugu Max is one model trained to route each task across a pool of open-weights and specialist models, including NVIDIA Nemotron, and to call instances of itself recursively, rather than a wrapper around a frontier model. It posts best-overall scores on six benchmarks including Terminal Bench 2.1, GPQAD and SWEFish at $2 per million input tokens and $6 out. The heavier Fugu Ultra v2 takes 74.3 on DeepSWE and tops SWEFish, beating models that cost three to five times more, and Sakana says it gets there without Fable 5.1 or GPT-6-Astra in its pool. Both are available now behind an OpenAI-compatible API, a one-parameter change for existing Fugu users.
Compressing your agent's terminal output saves tokens and does not save money.
Quesma ran Terminal-Bench 2.1 five times per task with and without RTK, 1,740 attempts across Claude Code on Fable 5.0 and OpenCode on DeepSeek V4 Pro. Fable came out 5% cheaper in aggregate but only 1% cheaper per task, while DeepSeek got 5% more expensive in aggregate and 17% more per task. Pass rates slipped both times, 84% to 83% and 71% to 69%. Terminal output was only 11% of Fable's input tokens and RTK touched just 31% of Claude Code's terminal calls, so the compressible share was small to start with and extra agent turns ate what it saved.
Your Claude Code plugin can now be scored against an eval suite and the score reproduced.
Version 2.1.269 adds claude plugin eval, which runs a plugin's own eval suite against Claude Code and returns the results as JSON and an HTML report. The same release raises the Workflow tool's per-run concurrent agent ceiling through CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS, settable from 1 to 256 for fan-outs that are inference-bound rather than CPU-bound. It also fixes two prompt-cache regressions: one after a response was cut off at the output-token limit and auto-resumed, another where resuming a session changed how much context was re-transmitted.
Three frontier models read Datasette's code and found subtle public-private data leaks.
Simon Willison, Sevban Dönmez and Alex Garcia audited Datasette with Claude Fable 5.1, GPT-5.6 and GPT-6 Astra, then spent close to a week on the fixes, which shipped as 1.0a39 and 0.65.4. The bugs hit instances that mix public and private tables, so anyone running Datasette on the public web should patch now. The team split the work so one person wrote tests demonstrating each issue while another implemented the fix, keeping two separate humans on every finding. Willison says audits by frontier models are going into all of their development work from here.
Rune's coding agent is an extension, not the editor, and the whole thing is now GPLv3.
Rune bundles code, terminals, CLI tools, language intelligence, debugging and AI agents into one composable multi-workspace environment, and its agent lives in cmd/rune-agent as an extension rather than in the core editor, a split the README frames as preserving manual programming alongside the automated kind. It is written in Go, licensed GPLv3 or later, and installs from a single curl of rune.build/install.sh, with DMGs for Apple Silicon and Intel on macOS 13.3 and later and x86_64 or arm64 Linux builds needing glibc 2.28 and OpenGL. Its help command opens Rune's own documentation as a workspace and starts an agent with those pages in scope.
