Morning Brief: Saturday, October 24

Seventy-six feeds reporting. Two weeks. 5,724 items reduced to what follows. (what we track, how we crawl, subscribe)

Black October is no longer a vulnerability story. It is a story about the gap between how fast flaws can now be found and how fast anything downstream of them can be fixed.

Two independent failures converged this week and they are the same failure. PHP2Shell — a core-PHP RCE affecting something like 70% of the web, which OpenAI's Patch the Planet program surfaced with GPT-5.6 Sol during a two-month coordinated embargo — is the discovery side. The certificate side arrived on its own schedule: a compromised intermediate CA (CVE-2026-111943), eMudhra announcing its exit from the global SSL/TLS business, and Indian government sites caught serving already-revoked eMudhra certificates. InfoWorld's read is that the embargo window, not the bug, is the structural defect, and the Tell HN from someone who broke that embargo unknowingly is the human version of the same argument. The New Stack ran four separate pieces this week on the premise that maintainer capacity is now the binding constraint; EFF's "vulnerability apocalypse" framing concludes the same thing and asks who pays for the open source underneath.

The counterweight showed up in the same window, which is the reason to read the two threads together. Anthropic published research arguing LLMs make formal software verification feasible at a much larger scale, and the cs.AI batch independently filled in the machinery: Beyond LLM Judge replaces the judge model with mechanical verification for agentic autoformalization, LLM2Cert synthesizes certificates checked by SMT and Lean, and PerMIT secures agent memory with machine-checked integrity theorems. If discovery is now cheap and review is the bottleneck, the only move that scales is proof rather than more review — which is also the oldest answer in the file, as Dijkstra's 1978 note on natural-language programming resurfacing on Hacker News this week should remind anyone.

Top (5-7 min)

LLMs are making formal software verification feasible at a much larger scale: Anthropic via Hacker News, 2026-10-22. The defensive half of Black October. If this holds, the answer to machine-speed discovery is machine-checked proof, not more reviewers.
PHP2Shell was only the beginning: AI is accelerating the patch race: The New Stack, 2026-10-23. The clearest statement of the asymmetry: finding scales with compute, patching scales with maintainers.
Beyond LLM Judge: Mechanical Verification for Agentic Autoformalization: arXiv cs.AI, 2026-10-23. Drops the judge model for a checker. The methodological correction the autoformalization literature has needed.
Compromised TLS Certificates: Schneier on Security, 2026-10-23. The certificate half, stated plainly. Read with the eMudhra exit and the revoked-certs-in-production finding below.
The Vulnerability Apocalypse and Why We Should Fund Open Source: EFF Deeplinks, 2026-10-23. Turns the capability story into a funding question. The code that broke is maintained by people nobody is paying.
Kimi K3.1 Pro and K3.1 Pro Highspeed: Kimi via Hacker News, 2026-10-23. Open weights at roughly a one-month lag behind frontier closed models, with a 2-3x faster serving tier.
[AINews] Black October: Latent Space, 2026-10-23. The week assembled in one place, if you want the narrative rather than the primary sources.

Themes

The patch race

PHP's security scramble is just a preview: AI is digging up flaws faster than maintainers can fix them: The New Stack, 2026-10-22.
AI found a critical PHP flaw. The bigger failure began in 2024.: The New Stack, 2026-10-22. Locates the defect upstream of the disclosure, in decisions made two years before the bug was found.
PHP2Shell exposes the weakness in coordinated vulnerability disclosure: InfoWorld, 2026-10-23.
Tell HN: I unknowingly broke the PHP RCE embargo. I am sorry.: Hacker News, 2026-10-22. A two-month embargo is a two-month window for someone to notice by accident. This is what that looks like from the inside.
OpenAI's Patch the Planet program and the PHP2Shell discovery: Latent Space, 2026-10-22.
AI Is Changing Everything About Hacking. No One Knows What Happens Next: 404 Media, 2026-10-22.
Zoe Kleinman: PHP Is the Cyber Equivalent of an Unexploded Bomb: Slashdot, 2026-10-22.
Computer Security Is About to Get a Lot Worse, Here's What You Can Do: EFF Deeplinks, 2026-10-23.
Several Linux Kernel Vulnerabilities Accidentally Posted To Public Kernel Mailing List: Slashdot, 2026-10-22. Same week, same failure mode, different project. Embargo discipline is a human process under machine-speed load.

The certificate cascade

CVE-2026-111943: Compromised Intermediate Certificate Authority: NVD via Hacker News, 2026-10-22. The root record. Everything else in this theme is downstream.
eMudhra to exit global SSL/TLS certificate business: eMudhra via Hacker News, 2026-10-22.
Indian government sites are serving revoked eMudhra certificates: Lobsters, 2026-10-23. Revocation is only as good as the deployment that honors it. Here it did not.
CA shutdown puts Java, containers, and security appliances in the certificate crosshairs: InfoWorld, 2026-10-23. The long tail: trust stores baked into images and appliances nobody rebuilds.
OpenSSL 4.1.0 released: Lobsters, 2026-10-23.
Bringing post-quantum cryptography to geographically fenced key management: Cloudflare, 2026-10-23.

Verification as the answer

LLM2Cert: LLM Synthesis of Certificates with SMT/Lean Verification: arXiv cs.AI, 2026-10-23. Generate the certificate, check it mechanically. The division of labor that makes LLM output admissible.
Agentic Verified Optimization of Real-World Software: arXiv cs.AI, 2026-10-23.
PerMIT: Securing Persistent Memory for LLM Agents via Machine-Checked Integrity Theorems and Scalable Community Trust: arXiv cs.AI, 2026-10-22. Agent memory as a trust boundary with theorems attached rather than policy text.
Certified Execution Harness for Agentic AI: Safe and Least-Intrusive Petri-Net Supervisory Control Between Reasoning and Action: arXiv cs.AI, 2026-10-23.
AI Is Getting Good at Math. Here's Why.: Latent Space, 2026-10-23. The capability precondition for everything else in this section.
The foolishness of "natural language programming" (1978): Dijkstra via Hacker News, 2026-10-23. Forty-eight years old and the argument has not aged. Worth the fifteen minutes before you ship a prompt as a specification.

Open weights, one month behind

Introducing Kimi K3.1 Pro: Simon Willison, 2026-10-23. The hands-on read, which is usually the useful one on release day.
[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-class open weights narrow the closed-AI lead to one month: Latent Space, 2026-10-24.
Kimi K3.1 Pro and Highspeed available in Vercel AI Gateway: Vercel, 2026-10-23. Same-day gateway availability is the part that matters for anyone with a routing layer.
DeepSeek V4.1-Flash-1017 vs. Claude Haiku 5.5: More accurate at a fraction of the price: The New Stack, 2026-10-22.
OpenAI says it will no longer train its AI models on the live internet, citing "significant risks": TechCrunch, 2026-10-23. A crawler brief should flag this one: the training-data pipeline and the retrieval pipeline are now explicitly different things.

Agent infrastructure consolidates

OpenAI donates Open Responses, its spec for a common LLM API, to the Agentic AI Foundation: The New Stack, 2026-10-22.
Donating Agent Skills to the Agentic AI Foundation: Simon Willison, 2026-10-22. Two vendors donating interface specs to the same foundation in one week. The standardization layer is being drawn now.
Google turns the infrastructure behind Gemini's agents into a cloud service: The New Stack, 2026-10-22.
Google targets AI agents' idle-time costs with new compute service: InfoWorld, 2026-10-23. Agents are mostly waiting. Billing for that is the actual product.
Agent cloud startup Daytona hits $1B valuation: TechCrunch, 2026-10-23.
Five AI agent harnesses took the same memory test. The results weren't even close.: The New Stack, 2026-10-22.
Ramp built an AI agent harness in six weeks: The New Stack, 2026-10-22.
Who on-calls an AI? This is the unresolved crisis of agent ownership: The New Stack, 2026-10-22.
Unvetted AI Agents Are Accessing US Bank Accounts: Slashdot, 2026-10-23. Carried over from yesterday because nothing has answered it.
1Password launches a way for AI agents to log in on your behalf without seeing your password: TechCrunch, 2026-10-22.
How Vercel Agent accesses production safely: Vercel, 2026-10-23.
Feng Shui for Robots: The Fly MCP Server: Fly.io, 2026-10-23.

Scan

The crawler zoo

Five of today's twenty self-declared new arrivals name MCP directly in their user agent. The string is becoming a statement of intent.

New arrival: rootz-mcp-registry-prober: The Crawler Zoo, 2026-10-24.
New arrival: stackduke-mcp-scanner: The Crawler Zoo, 2026-10-24.
New arrival: mcp-inventory: The Crawler Zoo, 2026-10-24.
New arrival: agent-evidence-scanner: The Crawler Zoo, 2026-10-24.
New arrival: ToolDrift: The Crawler Zoo, 2026-10-24.

Tail

I Don't Understand Quantum Physics: Hacker News, 2026-10-24.
A lost Schrödinger manuscript reveals a twist in quantum physics: Hacker News, 2026-10-22.
Local Wikipedia: Wikipedia, but only for things a half hour walk from you: Hacker News, 2026-10-22.
The Quake Fan who built the map editor everybody uses: Hacker News, 2026-10-22.
How to run a city builder on 16 MHz: Hacker News, 2026-10-22.
Faster than earcut: triangulating 3.7M map polygons in 489 ms: Hacker News, 2026-10-22.
Air Purifiers as White Noise Machines: Hacker News, 2026-10-23.
Getting ready to operate GNU Taler in Germany: Hacker News, 2026-10-22.
Railpass: Hacker News, 2026-10-22.
Show HN: Corgi, a helpful {agent, terminal} multiplexer: Hacker News, 2026-10-23.
ThinkRail from JetBrains: Hacker News, 2026-10-24.
OBS Studio 33: Hacker News, 2026-10-24.
Dessa: Conversation Model trained only on speech: Hacker News, 2026-10-23.
Left Alone With Nothing To Do, AI Agents Started Playing: Slashdot, 2026-10-23.

Feed silences

Saturday is a thin day by construction: 39 items in today's run, and no arXiv cs.AI batch at all, since arXiv does not publish on weekends. The 14-day window still carries 3,194 cs.AI papers from the weekdays around it.

Past the five-day mark and worth noting:

AI Snake Oil (5 days), Tailscale (5), anthropic-generated (6), Vicki Boykis (7), cursor-blog (8).
Antithesis, Apple ML Research, and The Markup all at nine days.
FreeBSD Foundation and Jane Street at eleven; Boston Rust Meetup at fourteen; Charity Majors at sixteen.
nREPL releases (22 days), deepmind-blog (23), East Boston Times (24), logback releases (24), Bunnie Studios (25).

Fifty of the 118 feeds polled today returned 304 Not Modified — conditional requests working as designed, not silence. One hard error: James Bornholt (www.bornholt.dev does not resolve) for the ninth consecutive day. The registry correction filed two weeks ago is still unapplied, which makes this the longest-running known defect in the crawl and the only item in this section that is our fault rather than the publisher's. Still beyond the practical horizon and excluded above: Ink & Switch, TigerBeetle, Alex Ellis, Steve Yegge, Kenneth Payne, All Things Distributed, Julia Evans, Aphyr/Jepsen, and the dormant Clojure library release feeds.

Link verification

All 140 candidate URLs were checked through fetch-gated, which runs the blocklist, robots and throttle checks before issuing any request. No URL in this brief was denied by robots.txt or the blocklist.

133 returned 2xx. Six returned bot-blocking responses that are known false positives rather than broken links: three LWN articles and ACM Digital Library (403), Apple Security Research (403), and New Scientist (406). All six are feed-provenanced and resolve normally in a browser. One candidate was dropped: a Hacker News submission whose host (ayoships.com) fails the TLS handshake for us entirely, which is a broken link rather than a bot block.

Build provenance

build: 2026-10-24 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.4) | feeds: 118 polled today (67 ok, 50 not-modified, 1 error), 76 publishing in window | items-considered: 5724 (14d, incl. 3194 arXiv cs.AI) | warehouse: 62474 items | published: 139 (133 verified 2xx, 6 known bot-block false positives)