Rust says LLMs can review, not create — and GitHub retires Spark
Rust's compiler team publishes the AI policy everyone else is arguing about, GitHub retires Spark, Anaconda buys an agent-security company, and yesterday's npm worm gets an independent recount.
Rust's compiler team just wrote down the rule the rest of us are still arguing about in Slack: an LLM can answer, analyze and review, but it does not get to be the thing that creates. GitHub retired Spark, its prompt-to-app builder, saying agents in your editor ate the use case — and gave Spark users until August 31 to export. Anaconda bought an AI-security company and published a number worth pausing on: 73% of the agent tool servers it scanned had vulnerabilities. And yesterday's npm worm got counted a second time, by someone else. The daily pulse of AI coding tools — what shipped, what matters, what's next.
Who's allowed to write the code
Rust says LLMs can review its compiler code, but not create it
Five teams in the Rust project have adopted an LLM policy for the rust-lang/rust monorepo — the compiler's own codebase — written up today by Jynn Nelson. The policy summarises itself in one line: "It's fine to use LLMs to answer questions, analyze, distill, refine, check, suggest, review. But not to create." What makes it worth reading rather than just linking is that it holds LLM-written code to a higher bar than human-written code, not a lower one: those pull requests "are required to have tests, full stop, regardless of how hard that is", and must stay away from soundness-critical work — the changes that decide whether the compiler can be trusted not to miscompile your program — unless the author is already a domain expert.
The stated reason is not code quality. It is review bandwidth and trust: there are 1,281 open pull requests on the repo, and as Nelson puts it, a polished PR no longer proves that a person understood it — or that a person is there at all. Pasting a reviewer's comment into an LLM and pasting the answer back is named in the policy as a breach of trust. Disclosure is mandatory for public LLM text, and lying about it carries a Code of Conduct penalty.
Two things the headline version of this drops, and you should keep: it is not a Rust-wide ban and not an official project stance on AI — it binds one repository. And there is a carve-out: pre-arranged, non-critical, well-tested LLM-created code is allowed with disclosure, under a clause marked explicitly experimental. The policy is candid that much of it is unenforceable; the goal, in its own words, is "to remove plausible deniability." Rust's leadership council is also weighing a dedicated LLM-policy sub-team, which would let the rules move faster than a thirty-person consensus allows. If you have been asked to write your team's AI policy this quarter, this is the most complete one a large open-source project has published. The Hacker News thread is worth the scroll too.
Aug. 5, 2026 · Jynn Nelson
What two vendors just killed and bought
GitHub retires Spark — export your apps by August 31
As of August 4, Spark — GitHub's build-an-app-from-a-prompt tool — stopped accepting new users and new apps. Anything already deployed keeps running, but if you want to keep editing what you built, you have until August 31 to export it to a repository. GitHub's own reason is the story: "AI models and agentic development tools have advanced significantly," and builders now do this work "through GitHub Copilot in the environments where they already work, including VS Code, Copilot CLI, and the GitHub Copilot app." That is a large vendor saying out loud that the standalone prompt-to-app surface did not survive contact with agents in the editor — worth a moment if you are currently evaluating anything in that category.
The trap for anyone with a Spark app: GitHub Models, the inference service behind Spark's llm() function, already retired on July 30 — those calls are dead now, not at the end of the month. Search your app for llm(); if it's there, you supply your own model provider, API key and billing from here. If it isn't, the Models retirement doesn't touch you.
Allison
Anaconda buys Enkrypt AI, which says 73% of agent tool servers have flaws
Anaconda announced on August 4 that it has acquired Enkrypt AI, which sells a security layer for agent stacks: pre-deployment red-teaming across "more than 300 attack categories," runtime guardrails against jailbreaks and data leakage, and compliance automation mapped to the NIST AI Risk Management Framework and the EU AI Act. The number that earns this a slot is Enkrypt's own: in the two months before the announcement it says it scanned more than 268,000 tools across 25,000 MCP servers and found more than 143,000 vulnerabilities, affecting 73% of those servers. MCP servers are the connectors that let an agent reach your real tools — your repo, your ticket tracker, your database — so that is a claim about the layer most people wire up once and stop thinking about.
Treat the figure as what it is: a vendor's count of a problem it sells the fix for, published by the company buying them, with no independent verification. It is still the only number anyone has put on the question. The stacking is deliberate and Anaconda says so directly — its "recent acquisition of Kilo Code extended that foundation into the agentic engineering environments", and Enkrypt now "brings the security, governance, and compliance layer that sits across all of it." Buy the coding agent in July, buy the thing that watches it in August.
David DeSanto
Yesterday's npm worm, counted again
JFrog counts yesterday's npm worm at 400+ packages
A follow-up to yesterday's lead, and the first thing to say is that this is the same incident, not a second campaign — same keyv compromise, same Shai-Hulud family. What is new is an independent count: JFrog puts it at 400+ packages across 1,700+ versions. That matters here specifically, because we published Aikido's figure of 868 before Aikido quietly revised it to 434 — and JFrog's independent tally lands with the corrected number, not the retracted one.
Two details worth carrying. On npm 12 or newer, preinstall hooks — the scripts a package can run automatically as it installs — do not run by default, so the payload does not fire there; that single line decides whether most readers need to do anything. And the persistence trick is aimed squarely at people who use coding agents: the worm commits .vscode/tasks.json, .claude/settings.json and three payload files to every branch it can reach, wired so that opening the repository in VS Code, or starting a Claude session inside it, runs the payload. The commits arrive titled chore: update config with a forged Co-authored-by: claude trailer. Claude did not write them; that line is camouflage.
Also worth your time
"Eight Myths on Software Engineering and GenAI" (ACM Queue) — six Microsoft and University of Victoria researchers on where the evidence and the narrative part company: developers spend roughly 14% of their time writing code, so a coding-only speedup has a low ceiling; AI-written lines of code is not a valid productivity measure; and one 2025 study found AI tools increased implementation time for experienced open-source developers by 18%. Published in May — it resurfaced yesterday and spent the day on the Hacker News front page (250 points, 204 comments), which is why it's here. Not news; still the most useful thing you can read this week if you are being asked to justify an AI rollout.
GitHub shipped two small Copilot changes on August 3 that are immediately usable if you drive Copilot from CI or from issues: you can now set the reasoning level for the Copilot cloud agent, and trigger Copilot automations from comments.
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.
The New Way is human-curated — a person picks every story. The summaries are written with AI (Claude) and reviewed before we hit send.
