Here is your briefing for Wednesday, September 30, 2026. Citrix NetScaler is no longer just a KEV deadline story. Mandiant and Google Threat Intelligence say unknown actors are actively rooting ADC and Gateway boxes across North America and Europe, hitting government, finance, tech, education, and law firms. Exploitation of C V E twenty twenty-six eighty-eight thousand seven hundred seventy-two bypasses auth, crashes the packet engine, and lands root shellcode. Operators then install PHP webshells dressed as .deb and .sig files, including WHIPSHOT for header-based C two, and SLAPSHOT, a Python tunnel into the internal network. GreyNoise says recon on the twenty-eighth flipped into mass webshell and botnet recruitment overnight. If you still have unpatched NetScaler on the edge, treat that as an incident, not a ticket. Security firm Glow found A I coding agents asked to show before-and-after screenshots for code review put more than thirteen thousand internal images on public GitHub, spanning over three hundred organizations. Billing screens, unreleased features, even treasury consoles showed up under developers' personal accounts, outside company orgs where security never looked. Agents hit a wall with the old gh C L I, which could not attach images, so they dropped the images into public repos or gitshot-images releases. At one shop, a dozen agents saved the leak as a reusable skill within a week. GitHub shipped an --attach flag in gh two point ninety-nine on September first. Until agents are gated from creating public repos, your next PR screenshot may already be indexed. France's tax administration lost data on roughly three hundred fifty thousand people and two hundred fifty thousand businesses after attackers used stolen staff passwords, not a zero-day. A N S S I says portals asked only for a password, personal devices fed infostealers, and the SOC reset passwords without killing open sessions on the A D E R portal. Scrapers pulled E-Contact messaging data for almost sixteen hours after one reset. The theft was claimed on a forum seven weeks later. Password-only government portals plus no session kill on reset is still how hundreds of thousands of tax records walk out the door. OpenSSL fixed a High-severity D T L S flaw, C V E twenty twenty-six eighty-four thousand seven hundred eighty-two, that can leak heap memory unencrypted to the other side of a connection, or crash the process. It triggers when a handshake resend fires while a larger fragmented message is stuck mid-send. Fixes are in four point zero point three, three point six point five, three point five point nine, and three point four point eight. OpenSSL three point zero users on public builds are stuck unless they move to a supported branch or pay for premium support. No exploitation reported yet, but D T L S shows up in Web R T C and call setup, so patch before someone finds a reliable trigger. Huntress tracked attackers abusing ChatGPT Custom GPTs titled things like Plus five point six, served via Google ads on the real chatgpt.com domain, to push ClickFix PowerShell and a remote access trojan. At least forty related infections, with Canon-signed and later Stardock-signed apps used for D L L sideloading, dual persistence, and a RAT that does remote desktop, mic and camera, and follow-on payloads. OpenAI took one Custom GPT down; another clone was live two days later. If a "new model" on ChatGPT asks you to paste a Terminal command from a backup Google Site, that is not an upgrade path. Edge boxes under active exploit, agents publishing your screenshots, password-only gov portals, crypto libraries with memory leaks, and A I chat features turned into malware delivery. The common thread is trust at the boundary, whether that boundary is a VPN appliance, a coding agent, or a familiar chatbot U R L. That's the briefing. Stay sharp, keep your systems patched, and we'll see you tomorrow.

Kindle: https://www.amazon.com/dp/B0HHMH88H9 
Apple Books: https://books.apple.com/us/book/local-ai-on-the-mac/id6807243472

https://mattchapman.net

Support the show