Watch on YouTube · subscribe for the daily brief

Firefox publishes security fixes disguised as ordinary code changes, to slow attackers down until the release reaches users. Epoch AI reports that GLM 5.3, an open-weight model from Z.ai, saw through that and built working attacks against six Firefox releases at about $1,000 each.

What Epoch tested, and what it did not

• The target: every Firefox release from September 17 to October 26: 156.0, 156.0.1, 157.0, 157.0.1, 158.0 and 158.0.1 (Unite.AI).
• The setup: meant to approximate a low-skill attacker. Basic tools, the target version, and only information public on each release day.
• The result: a web page that, once visited, reads a file from the visitor's computer and sends it out, with no other interaction.
• The time: up to 48 hours per target allowed; the fastest attempt took 12 hours. Some headlines read this as "under 48 hours", which is the limit, not the result.
• The safeguards: none triggered at any provider, including Z.ai.
• The bonus bug: in every compromise the model also found an unpublished vulnerability. Mozilla patched it in Firefox 159.0 on October 27.
• The limits: this was a lab test. Similar attacks on Chrome have not succeeded so far, and Epoch does not treat that as strong evidence that they cannot.

What to do: update to Firefox 159. If you maintain open-source software, treat the time between a public fix and your release as exposure, and shorten it. Epoch suspects closed-source projects are exposed too when a model can read upcoming changes.

A Linux flaw with two scores: 9.8 or 5.9

CVE-2026-98432 is a use-after-free in the kernel's IPv4 TCP code, going back to kernel 4.4. The kernel CVE team scored it 9.8. Red Hat scored it 5.9, calling it a potential crash that is complex to exploit. Anthropic researchers then reported a remote root exploit, written by an unreleased model in about 53 minutes, against a server with a Multipath TCP listener.

The catch is Go: since 1.24 it enables MPTCP by default for TCP listeners on Linux. Ubuntu kernel fixes were still in progress on October 30; Debian has shipped them. Disabling MPTCP alone is not enough, because ordinary TCP can still trigger the bug. No exploitation in the wild is confirmed.

The Claude Code extension on Cursor

Backdoored versions 2.1.312 and 2.1.313 sat on Cursor's marketplace for about 11 hours on October 29, reaching at most 5,000 users. Phorion says they were posted to the official Anthropic account; Anthropic says its systems were not breached and the actor used access to Cursor's marketplace. Nobody has explained the way in. If you used Cursor that day, look for a key labeled `svc-telemetry` in `~/.ssh/authorized_keys`, and follow Cursor's steps. Uninstalling may not be enough.

Also today

• Intel: down around 28% on Friday, about $12 billion off the US government's stake. It blames a PC slowdown from memory prices and availability, a direct cost of AI demand. The stock is still up more than 70% this year.
• Google Suncatcher: the prototype is online and wrote a haiku with Gemma 4. The useful part is the constraint: at up to 1 kW it can run one of its four TPUs for only about 20 minutes at a time.
• Reddit: moving to logged-in-only, testing outside the US this quarter, after losing over a million search referrals a week in October. Huffman on Google's AI Mode: "AI Mode does not drive traffic."
• AI training opt-out: Google, Microsoft, OpenAI and Anthropic back a draft that blocks training scrapers with one robots.txt entry. A promise to honor a text file, and still a draft.
• Postgres 19: likely November or later, says committer Robert Haas. Back-patched commits rose from 724 to 1,169 in the same period, because LLMs made bugs easier to find than to fix.
• Anthropic credit: Blackstone now expects a high junk rating after the IPO, per the FT. It is a forecast.
• DoorDash: outside AI agents have placed maybe several hundred orders in total, against nearly a billion orders last quarter. The connector is under a month old and limited to beta partners.
• Ghost students: the DOJ charged five people and counts over 830,000 potentially fraudulent aid accounts, with $32 million in actual loss tied to these suspects so far.

The through-line

Finding the flaw got cheap. Everything after it did not. Postgres cannot fix bugs as fast as it now finds them, and Firefox users wait for a patch that is already readable. Today: update Firefox to 159, and check whether your Go services on Linux have an MPTCP listener.

Sources covered

1 story · 90 sources (64 news, 23 company, 3 voices)

1. Tell HN: Scribd Breached
   📰Hacker News↗🏢AWS AI/ML↗📰SiliconANGLE↗🎙️Simon Willison↗🏢AWS AI/ML↗📰TechCrunch↗🏢Databricks↗📰SiliconANGLE↗🏢Databricks↗🏢Databricks↗📰TechCrunch↗📰The Hackers News↗📰TechCrunch↗📰Wired↗📰The Verge↗🏢Starburst↗🏢Hugging Face↗📰TechCrunch↗📰TechCrunch↗📰Ars Technica↗🏢AWS AI/ML↗🏢AWS AI/ML↗📰TechCrunch↗📰TechCrunch↗📰Ars Technica↗📰BleepingComputer↗🎙️Dwarkesh Patel↗📰The Verge↗📰Ars Technica↗📰The Verge↗📰The New Stack↗🏢Google Cloud AI↗📰The New Stack↗📰The Decoder↗📰The New Stack↗📰Ars Technica↗🎙️Damien Benveniste↗🏢NVIDIA↗📰The Verge↗📰The Decoder↗📰TechCrunch↗📰The Decoder↗📰Ars Technica↗📰The Verge↗📰The New Stack↗📰The Decoder↗🏢Microsoft Research↗📰TechCrunch↗🏢Vercel AI↗📰Ars Technica↗📰LWN.net↗📰TechCrunch↗📰Hacker News↗📰LWN.net↗📰SiliconANGLE↗📰SiliconANGLE↗📰The New Stack↗📰The Verge↗📰SiliconANGLE↗📰BleepingComputer↗📰TechCrunch↗📰TechCrunch↗📰InfoQ↗🏢OpenAI↗📰The Verge↗📰The Verge↗📰The New Stack↗🏢OpenAI↗📰The Decoder↗📰SiliconANGLE↗📰SiliconANGLE↗📰SiliconANGLE↗📰InfoQ↗📰Hacker News↗📰The Verge↗📰InfoQ↗🏢Supabase↗🏢Supabase↗📰Hacker News↗📰The Verge↗📰InfoQ↗📰InfoQ↗🏢Vercel AI↗🏢Vercel AI↗📰Hacker News↗📰Wired↗🏢Temporal↗🏢Vercel AI↗🏢Vercel AI↗🏢OpenAI↗