The story that will keep most of us busy this week isn’t about AI: PHP2Shell hits a huge chunk of the web and is already being exploited. If you run anything on PHP, start there.
On the AI side, the most interesting thing for developers is seeing models do real security work, on both sides of the fence.
Security
PHP2Shell: two critical bugs allow unauthenticated code execution on PHP servers
CVE-2026-109618 and CVE-2026-109619 allow remote code execution and affect WordPress, Drupal, Joomla and Magento. The fix shipped early after a leak broke the embargo, and attacks have already started. The NSA has issued an advisory urging people to patch.
Why it matters: If you maintain any PHP application, update to the fixed release today: exploiting it needs no credentials.
Source: PHP Foundation
Security
PAN-OS GlobalProtect (CVE-2026-0257) exploited by ransomware groups
Unit 42 confirms active exploitation of this GlobalProtect vulnerability by the Qilin and Inc groups. The same week saw exploitation of Plex Media Server (Plex2Shell) and Siemens Software Center.
Why it matters: The corporate VPN is ransomware's favourite way in; if your company runs GlobalProtect, make sure it's patched.
Source: Unit 42
Security
Storm-0940 uses an agentic pentesting system in a supply chain attack
Microsoft details how a China-linked group used ARTEX, an AI agent based pentesting system, to compromise the ConnectWise supply chain.
Why it matters: Attackers already automate recon and exploitation with agents, so the gap between a bug going public and you getting hit keeps shrinking.
Source: Microsoft Security Blog
AI
Claude checks security properties in 123 open source libraries with formal verification
Anthropic published research in which Claude formally checked security properties in 123 open source libraries and found bugs that had gone unnoticed for a long time.
Why it matters: It hints at where code review tooling is heading: past linters, towards AI-assisted formal checks.
Source: Anthropic
Cloud
OpenAI models and Codex come to Google Cloud
Google Cloud customers will be able to use OpenAI's frontier models and Codex from Model Garden, with general availability in the coming weeks.
Why it matters: If you're on Google Cloud, you'll get OpenAI models with the same billing, IAM and networking, without onboarding another vendor.
Source: OpenAI
AI
Mistral and NVIDIA team up on AI for industrial engineering
The partnership connects Mistral's models with NVIDIA's simulation tools, and they plan benchmarks for agents working on chip design.
Why it matters: Agents are moving beyond code and office work into simulation and hardware design, with benchmarks of their own.
Source: Mistral AI
Web
Meta's new terms allow training AI on business websites
From October 30, Meta's commercial terms allow content from business websites to be used for AI training without a separate opt-out. Meta says robots.txt can limit access.
Why it matters: If you run a site and don't want it used to train models, check your robots.txt before October 30.
Source: Nieman Lab
Tools
Vercel: customers' AI spend grew 13x in nine months
According to its AI Gateway report, customers spent about 13 times more on AI in September than in December, even though the unit price fell 47%.
Why it matters: Cheaper tokens don't mean a smaller bill: set limits and track AI usage in production.
Source: Vercel
If any of this made you change something in your stack this week, I’d love to hear about it in the blog comments.