My security topic of the day ties together everything else I posted today. I'm a few days late on this report, but it's the context for why the vCenter and Ruby posts say "now" instead of "this quarter."
Google Threat Intelligence Group published its latest AI Threat Tracker on November 16. A lot of it is the usual list of tools and actors. One finding is worth slowing down for.
What Google Actually Found
Between July and September, GTIG found an exposed staging server running a multi-agent framework called exploit_agent. Per GTIG and TechTarget's write-up, it:
- continuously compares patched and unpatched versions of popular web applications (patch diffing),
- drafts exploits for whatever the patch fixed,
- runs them against targets, then uses scripts for recon, brute-forcing and installing webshells (and removing other attackers' webshells, which is a nice touch),
- does it without a human in the loop.
GTIG says it's the first time it has seen LLM frameworks fully automate finding and exploiting bugs like this. The operator was going after French media, entertainment and education websites. Not exactly nation-state crown jewels: just whatever was unpatched.
The report also walks through the PaperCut zero-days from August. Huntress had working exploits within a day of disclosure, and GreyNoise saw a campaign four days after disclosure.
Let's Keep This in Proportion
This is not "AI can now hack anything." Patch diffing is decades old. Skilled humans have always been able to turn a patch into an exploit; the bottleneck was how many skilled humans there were. What changed is that the bottleneck is getting cheaper, so more patches get turned into exploits, faster, by less skilled people.
The other thing GTIG noted is just as useful: after the PHP2Shell disclosure, "hundreds" of claimed proof-of-concepts showed up that didn't work. Automation produces a lot of noise too. That cuts both ways for defenders: more real exploits, and a lot more junk to sort through.
The practical takeaway isn't fear. It's math. If the time from patch to working exploit is moving from weeks toward a day, a 30-day patch SLA for internet-facing systems is a policy written for a different decade.
What To Do
- Tier your patch SLA by exposure, not just CVSS. Internet-facing, unauthenticated, and on KEV or with a public PoC: days, ideally hours. Internal and authenticated: the normal cycle is fine.
- Shrink what's internet-facing. The fastest patch is the one you don't need on day one because the app isn't reachable from the internet. Admin consoles, management planes and internal tools belong behind authenticated, brokered access. (I work at Zscaler, so yes, this is my bias, and it's still true whoever you buy it from.)
- Know your inventory before the advisory lands. You can't patch in hours if it takes you two days to figure out where Ruby 3.3 or vCenter 8.0 runs.
- Pre-stage emergency change. Get the "we patch critical exposed systems same-day" approval signed off now, not during the incident.
- Watch for webshells after you patch. Automated exploitation tends to drop a webshell first and ask questions never. Patching doesn't remove one that's already there.
- Lock down your own AI keys. The same report covers LLMjacking: stolen model API keys running up huge bills within hours. Keep keys out of repos, scope them, set spending alerts.
Bottom Line
Attackers automated the boring part: reading patches and writing exploits. The answer is to automate our boring part: knowing what we run, hiding what doesn't need to be public, and patching the exposed stuff fast. Same fundamentals, shorter clock.
Sources: Google Threat Intelligence: AI Threat Tracker, TechTarget