Skip to content
subscribe --daily

cat newsletter.md

One email each morning

Every story published in the previous twenty-four hours, with the same opening summary that appears on the site. No roundups of other people's writing, no sponsored items, nothing that was not published here that day.

What you are agreeing to

  • One email a day, at 07:30 IST — and only when something was published in the preceding day. A quiet day means no email rather than a filler one.
  • Titles and summaries, not full articles. The email is a list; the writing stays on the site.
  • Confirmation first. Subscribing sends one email with a link. Nothing else arrives until that link is clicked, so an address typed by somebody else never ends up on the list.
  • One click to leave. Every email carries an unsubscribe link that works without a login or a reply.
  • The address is used for this and nothing else. It is not sold, shared or used to advertise. See the privacy policy.

What it looks like

The stories from 2026-11-13, in the order the digest sends them. This is the email, not an illustration of one.

Subject: Root Notes — 6 new stories today

Root Notes

Breaches, patches and the infrastructure underneath — in five minutes.

  • A scanner reported it eight days before the humans did

    Apache's advisory for a critical HTTP/2 flaw credits two independent finders. Anthropic's scanner reported it on 26 October. Google's vulnerability research team reported the same bug on 3 November.

  • Thirty-five thousand apps, public because a subdomain is not a secret

    Researchers counted 37,571 applications built with Grok's app builder and found 35,578 of them open without a login. The access control was a random subdomain, and subdomains are published on purpose.

  • A model found the way to backdoor models

    Microsoft Research showed that rewriting fewer than ten dormant neurons turns an open-weight model into a sleeper backdoor, with no retraining. The post also says the work began with Claude Code, which discovered most of it autonomously.

  • Seven hundred and forty thousand dollars of somebody else's inference

    Railway's incident report puts the unauthorised spend on stolen customer AI keys at about 740,000 dollars. It says AI provider keys were the primary target, and that it will cover from its own balances what the providers will not refund.

  • The advisory says a model found it

    Three days ago we wrote that Mozilla's advisory never records when a model found a bug. This week Microsoft's acknowledgement names the tool, and Apache's gives the scanner its own tracking identifier.

  • The counting spread tripled in a month

    In October four organisations counted the same Patch Tuesday and landed 54 CVEs apart. In November five counted it and landed 169 apart, with four of them clustered and one sitting 165 above the rest.

Unsubscribe

subscribe --daily

One email each morning. No filler.

Would rather not hand over an address? The same stories go out on RSS, and all 506 of them are in the archive.