powering productive workplaces
latest
Front page
FeatureTrust & Risk18m · 01:34 GMT · 5 min read

Google: AI Agents Now Turn Patches Into Exploits With No Human Help

Google's threat intelligence team says it has caught attackers using AI agents to turn security patches into working exploits with no human involved. Here's what the report actually shows, and what it means for the meeting platforms, messaging tools and AI agents your teams rely on every day.

UC Today Developing Story artwork

Google says it's seen something security people have been bracing for: an attack pipeline where AI agents find the bug, write the exploit and fire it, with no human signing off on any step.

That's the headline finding in the latest GTIG AI Threat Tracker, published today (November 16) by Google Threat Intelligence Group and Mandiant. Google's own summary on The Keyword doesn't hedge. It's titled "The era of autonomous attacks has arrived."

Here's what the report actually says, and why anyone running a collaboration estate should care before Microsoft Ignite opens this week.

What Google Found

The centerpiece is a framework GTIG has dubbed exploit_agent. Researchers found it on an exposed staging server. In the report's words:

"In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time."

The framework plugs into multiple frontier and local large language models and runs "continuous differential analysis of popular web application software." Put simply, it watches for patches, compares the code before and after, works out what was fixed, drafts an exploit and runs it. GTIG says it "automatically executes them without human intervention."

The server also held data stolen from victims, plus scripts for reconnaissance, brute-forcing and installing webshells. Comments in the configuration files were written in Russian. Victim telemetry suggested the operator was interested in French websites in media, entertainment and education.

Google doesn't name a group or a country behind it. It's also worth being precise about scope. The "first time" claim covers fully automating discovery and exploitation, and the targets were web apps, not a big-name enterprise platform.

The Patch Window Is Shrinking

This is the part I'd underline for IT leaders. GTIG says attackers are using LLMs to "rapidly weaponize n-day vulnerabilities," meaning flaws that already have a fix. The report says that is "shrinking the patch window for organizations."

The numbers back that up. Google says disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits rose from a 2025 average of eight a month to 11 a month in 2026, then spiked to 22 in August.

There's a second sting. GTIG says the lower barrier to entry "has fueled a wave of LLM-generated exploits that are claimed to work but don't." After PHP patched two flaws on October 21, hundreds of claimed proofs of concept appeared, and many didn't work.

Beyond exploit_agent, GTIG says threat actors are using open source multi-agent offensive frameworks, including ARTEX, CAIRN, CyberStrikeAI and PentAGI. It saw CAIRN used against targets including the US aerospace sector.

Why This Lands in the Collaboration Stack

Here's my analysis. Collaboration platforms are internet-facing, always on and patched on a cadence that often depends on change windows, room systems and nervous business owners. Meeting clients, web conferencing gateways, session border controllers and contact center front ends all fit that description. If the time from patch to working exploit is heading toward hours, a monthly window starts to look like an open invitation.

The social engineering findings land even closer to home. GTIG describes a suspected North Korea-linked cluster it calls MIDNIGHT NEPTUNE. It built fake recruiter profiles with AI-generated images, then steered targets to "fraudulent meeting platforms" carrying ClickFix lures. Once inside, the group tampered with AI coding assistant settings so a backdoor downloaded at startup. It even named GitHub access tokens after Microsoft Copilot to blend in.

We've covered this pattern before, with bogus Zoom, Teams and Meet invites targeting busy employees. What's new is the AI layer on top: synthetic personas, LLM-written lures and AI tools inside the company used as the way in. GTIG also says North Korean IT workers continue to use LLMs to write resumes and answer job screening questions while posing as Western candidates.

So the meeting invite, the recruiter message and the friendly agent in your IDE are all attack surface now.

Speed Over Stealth, For Now

There is some good news. GTIG says these frameworks "currently trade stealth for velocity." Unrestricted autonomous execution is noisy and leaves traces. Google expects that to change as the tools mature.

Florian Roth, head of research at Nextron Systems, made a point on X today that I think buyers should hear. He argues the industry is treating ARTEX and CAIRN like old-style command-and-control kits, and that's wrong.

"There is no special defense against ARTEX. Tomorrow the same attacker could use a different framework, a different model and completely different tools. The names are disposable. The underlying attack techniques are what we need to detect."

Roth's post is a useful check on any vendor this week promising to "detect agentic attacks." Ask what behavior they detect, not which tool names they know.

Google, naturally, has a product answer. It's pitching Google AI Threat Defense, which it says combines Gemini and other frontier models, Wiz's risk prioritization, CodeMender's automated patching and Mandiant's intelligence. I'd treat that as a vendor claim to test, not a conclusion.

What I'd Do Now

Measure your real patch window. Ask how long it actually takes from a vendor fix to deployment on your internet-facing UC and contact center components. Then ask whether that number survives an attacker who works at machine speed.

Press your UC vendors. Ask how fast they ship fixes for cloud services, who patches the hosted pieces, and how they'll tell you when an n-day is being exploited.

Inventory the agents. Find out which AI assistants and coding agents run inside your estate, what they can execute and who controls their configuration files.

Retrain for fake meetings and fake recruiters. "Your audio isn't working, run this fix" should set off alarms. So should a recruiter who only wants to talk on a platform you've never heard of.

Detect behavior, not brands. Take Roth's advice. Signatures for a named framework won't save you when the attacker swaps it out next week.

I'll be at Ignite this week, and I'll be asking vendors how they're thinking about all of this. If you're seeing AI-driven attacks against your communications estate, I'd like to hear about it.

rate this story
helps rank stories across uc today
Join the conversation

Does this reflect your experience?

opening the room…