AI is flooding the world with vulnerability reports, and the patching scramble that follows is starting to cause outages of its own. The harder question is how many of those bugs actually matter.

1. Report: AI agents found and used a zero-day on their own

The Register’s headline is blunt: attackers in China deployed AI agents that autonomously found and exploited a zero-day to take over a university server. The text we could retrieve was only that headline, so details such as how long the agents ran, what tools they used, and how the attribution was made are not something we can verify.

Still, the claim is worth taking seriously. Exploiting a known bug quickly is one thing. Finding an unknown one and using it without a human steering is another, because it removes the gap defenders count on between a flaw existing and someone weaponizing it.

If the report holds up, the university server is a modest target but a telling one. Attackers rarely debut tools on their best victims. The open question is whether anyone publishes enough detail for defenders to test their own detection against this kind of behavior.

The Register

2. 100,000 bugs, 300 exploited: the outage story is about patch speed

The 2026 vulnerability count passed 100,000 this week, per WIRED, against roughly 25,000 in 2022. But only 300 have made CISA’s known-exploited catalog. VulnCheck counts 1,500, about 1.5 percent of the total. Security researcher Joe Slowik says the number of bugs you need to care about has stayed roughly constant. What changed is how fast they get weaponized.

That speed is the pressure point. WIRED says Microsoft’s 365 outage came from a flawed update, Apple blamed an internal update, and officials blamed the NSA network outage on botched patching after October’s PHP fixes. WIRED stresses there is no evidence of cyberattacks or a coordinated AI campaign, and each incident looks routine on its own.

Satya Nadella says a patch now has to go from available to deployed instantly when it’s being exploited. Slowik says those turnaround times defeat patch verification. Cloudflare’s Matthew Prince expects two years of sky-is-falling before AI helps defenders catch up.

The AWS and FAA items came to us as headlines only, so we can’t tie them to patching. Researcher Lea Schönherr makes the sharpest point: responsibility lies with whoever gives AI access to infrastructure.

The Register · WIRED · Ars Technica

3. A one-byte Ruby bug reaches default Rails apps before login

The German letter ß uppercases to SS, two bytes where Ruby’s legacy single-byte encodings budgeted one. Per Cloudflare, citing Veria Labs, a string of 21 of them overflows a heap buffer by a byte. That is ruby2shell (CVE-2026-132825), rated Critical. Rails and classic-style Sinatra reach it by default through Rack’s _method override, which upcases input during routing, before authentication. Fixes are in Ruby 3.3.14, 3.4.13 and 4.0.10, and Ruby says proofs of concept achieving code execution exist.

PHP’s phpAfterShell (CVE-2026-126714, CVSS 9.2) is messier. Exploitation needs particular session conditions and a suitable deserialization chain. Cloudflare says none of the 17 PHP flaws patched this week appeared in CISA’s exploited-vulnerabilities list as of November 19. Its post also implies PHP 8.6 tightens a session default, since an explicit upload-progress setting would override the new one.

Cloudflare sells the WAF it recommends, but it concedes the WAF is a compensating control. Ruby bundled in GitLab, Mastodon and Discourse is the likelier trouble spot: inventorying runtimes is slower than flipping a rule.

Cloudflare Blog · Cloudflare Blog · The Register

4. Meta rethinks its 2030 climate goal as geothermal remains a bet

Ars Technica reports that Meta is re-evaluating its 2030 sustainability goals amid record emissions from AI data centers. That is not an abandonment, and the text we retrieved was only the headline and a one-line teaser. But a company with Meta’s money reopening a climate target suggests clean power isn’t arriving on the schedule AI buildouts demand.

TechCrunch’s geothermal roundup shows where hopes are pinned. Quaise says Google will pay more than $500 million to secure 50 megawatts from an Oregon project tapping heat three miles down. Mazama says a CO2-cooling technique cut drilling time another 20 percent. Drilling is a quarter to a third of a plant’s cost, so fewer, hotter wells could change the economics.

The catch is timing. Those are company claims about projects still to be built, while data-center demand is already here. Google paying a premium to lock in early supply is also a reminder that founders may face a power queue as well as a chip queue.

Ars Technica · TechCrunch

5. Ars says 2 million planted canaries exposed a leaky AI supply chain

Ars Technica’s headline says a researcher planted 2 million canaries into the AI supply chain and a privacy disaster surfaced. Its teaser says prompts with unique, trackable secrets prove that data doesn’t stay private. We could only retrieve those lines, so where the leaks appeared and which vendors were involved remain unknown to us.

The method is the useful part. A canary is a unique string that should appear nowhere else, so finding it elsewhere is hard to explain away. That moves the privacy debate from vendor assurances to measurable evidence.

The Register’s report that Signal is blocking prying AI agents points the same way, though we only have its headline. If you send customer data through AI tooling, ask vendors where it goes after the API call, not just whether they train on it.

Ars Technica · The Register