Laude's agent Audel fixed its own bug unprompted and Headlong brings always-on agents. Plus a real, patched RCE in vLLM's tool-call parser, a screenshot-free daily memory for Claude Code, and FSFE on AI code copyright.
The Laude Institute and MIT open-sourced Headlong, a bash-based framework for agents that think continuously instead of waking only per request, and say their own agent fixed a bug in itself without being asked. A security researcher traced a real, already-patched remote-code-execution bug in vLLM's Qwen3 Coder tool-call parser into a broader warning: an inference server's own parsing code is attack surface a model can reach, not just the sandbox wrapped around it. A solo builder shipped a macOS tool that feeds Claude Code a plain-text record of your day, built from the accessibility API instead of screenshots. And the Free Software Foundation Europe published a legal read that most vibe-coded software may not be copyrightable at all, which means it can't carry the license someone attached to it. Four items this pass share one throughline: how much a coding agent can reach without anyone watching.
What an agent can reach #
Laude Institute and MIT launch Headlong, open-source agents that think continuously
The Laude Institute and MIT open-sourced Headlong, an agent "microharness" in under 10,000 lines of Bash: instead of waking only per request, its agents keep a continuous internal thought stream and decide for themselves when to act. Laude says its own agent, Audel, fixed a bug in its own code unprompted, audited a teammate's branches, and has had 50-plus commits pulled into main. Background operation runs roughly $1 to $2 an hour. That's Laude's own account, nobody else's yet.
A patched vLLM bug let an LLM execute code on its own host
Security researcher Boyd Kane argues an LLM could gain control of the machine serving it, not just its sandbox, and points to a real, patched example: vLLM's tool-call parser fed model-generated arguments into Python's eval()
, "allowing the LLM to execute arbitrary code on the host machine." That bug, CVE-2025-9141, scored 8.8 on CVSS and is fixed in 0.10.1.1. Kane rates a model finding one on its own "somewhat likely," and isn't sure. Self-host an inference server and its parser is attack surface.
Built this week, and the law around it #
Ambient Context feeds Claude Code your day, built with no screenshots
A macOS menu-bar tool called Ambient Context reads the text of your focused window every few seconds through the Accessibility API, no screenshots, and writes it to a local daily markdown file. Point Claude Code at the folder and it can answer "what did I work on Tuesday?" Nothing leaves the machine; passwords and card numbers are scrubbed before writing. One commenter who tried the same approach earlier this year called it "barely sufficient" because most apps aren't wired up right for it.
FSFE says fully AI-written code can't be copyrighted or licensed
The Free Software Foundation Europe published a legal analysis: code an LLM writes entirely on its own isn't copyrightable, so it defaults to the public domain and can't carry a Free Software license. "Only a natural person can be a copyright owner," it notes, flagging the UK and Ireland as outliers. Code mixing substantial human authorship can still qualify. FSFE's advice is to structure your own program rather than ask a model for the whole thing, and disclose which system helped.
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
Also worth your time #
A Windows veteran vibe-coded a Task Manager clone from a 107-page spec, now on Mac and Linux— Dave Plummer, who wrote the original Windows Task Manager, fed Claude Code a 107-page spec and had a rough app running before his son left the hospital after a procedure; it now animates at 60Hz and runs on Windows, macOS and Linux.
The New Way is human-curated — a person picks every story. The summaries are written with AI (Claude) and reviewed before we hit send.