The Human Just Left the Attack
For years, the security industry has warned that AI would one day run cyberattacks on its own. On November 17, 2026, Google said that day has arrived.
Google Threat Intelligence Group (GTIG) reported that, in the past quarter, it saw threat actors use AI frameworks to discover and exploit vulnerabilities without a human in the loop, for the first time. GTIG calls it "a paradigm shift towards intrusions occurring at machine speed."
The attacks themselves are not new. What changed is who runs them, and how fast.
Inside exploit_agent
The clearest case came from an exposed staging server. On it, GTIG found a framework its operators called "exploit_agent": a multi-agent system connected to several frontier and local AI models.
Its workflow is short:
| Step | What the agents do |
|---|---|
| 1. Diff | Compare a popular web application's code before and after a security patch is announced. |
| 2. Find | Identify the vulnerability the patch was written to fix. |
| 3. Weaponize | Draft a working exploit for that vulnerability. |
| 4. Launch | Run the exploit against real targets, with no human approval in between. |
The same server held stolen victim data, brute-force scripts and tools to install and remove webshells. Configuration comments were written in Russian, and one label read "for auto-exploitation (exploit_agent pilot)." Victim telemetry pointed to French websites in media, entertainment and education.
This technique has a name: patch diffing. Human researchers have done it for decades. It is slow, skilled work. It is also exactly the kind of task large language models do well: read two versions of code, explain what changed, and write code that uses the difference.
Not One Tool, an Ecosystem
exploit_agent is not alone. GTIG and Mandiant saw attackers using several AI offensive frameworks, some of them open source, and some groups building their own.
| Framework | What was reported |
|---|---|
| ARTEX | The most used, mainly by Chinese-speaking groups. The FBI issued a FLASH alert on November 13, 2026: in use since at least August 2026 against financial services, government, technology, defense, healthcare, energy and other sectors. |
| CAIRN | Used in activity targeting the US aerospace sector. |
| PentAGI | Used for systematic reconnaissance and vulnerability scanning across organizations in Asia. |
| CyberStrikeAI, HexStrike AI | Also observed in attacker operations. |
| Custom frameworks | GTIG saw a China-nexus group building its own automated attack framework, likely for espionage. |
According to the reporting, these tools can find and exploit simple vulnerabilities and adapt their tactics based on how a target responds.
The Patch Window Is Closing
Most vulnerability management programs rest on one quiet assumption: after a patch is released, defenders have time. Attackers need days or weeks to reverse engineer the fix and build a reliable exploit. Defenders patch in between.
AI breaks that assumption. GTIG notes that automating vulnerability analysis and exploit development lets threat actors shorten the time between a patch's release and its active exploitation. The broader numbers point the same way: GTIG's September analysis found that vulnerability disclosures and exploits nearly doubled between January and August 2026.
There is a side effect too. After patches for two PHP flaws (CVE-2026-109618 and CVE-2026-109619) landed on October 21, hundreds of non-functional proofs of concept appeared online. The security community had to check each one. AI does not only speed up attackers. It also floods defenders with noise.
LLMjacking: Your AI Bill as the Target
The same reporting describes a second trend: attackers stealing access to AI itself. Mandiant responded to several LLMjacking incidents in 2026, where attackers hijacked cloud environments to consume the victim's compute and AI model quotas.
- In some cases, fraudulent charges passed $100,000 in a single day.
- In the worst cases, the total reached $1 million before the abuse was stopped.
- One botnet using leaked credentials generated millions of AI API calls in about 24 hours before the keys were revoked.
- Stolen access is resold through "LLM Mixers," at $0.10 to $25 per million tokens.
For any company running AI workloads in the cloud, an exposed API key is no longer just a data risk. It is a direct financial risk.
The Good News
The Attacks Still Need the Same Doors
GTIG stresses that the mechanics of intrusions have not changed. These agents still abuse unpatched systems and compromised credentials. Every control that closes those doors still works against an AI attacker.
They Trade Stealth for Velocity
According to GTIG, these frameworks generate traffic and code traces that make them easier to detect. An agent that tests exploits at machine speed leaves a loud trail. Teams that watch their logs have a real chance to catch it.
What to Do Now
| Action | Why |
|---|---|
| 1. Patch internet-facing systems in days | Prioritize anything exposed to the internet. The window between a fix and a working exploit is shrinking fast. |
| 2. Watch the days after a patch | Expect bursts of exploit attempts right after a vendor releases a fix. Raise monitoring on affected systems until they are patched. |
| 3. Lock down credentials | Enforce MFA, remove unused keys, and rotate any exposed cloud or AI API key. Set budget alerts so LLMjacking shows up as a cost spike, not a surprise invoice. |
| 4. Automate detection and response | Machine-speed attacks need machine-speed defense. Automate the first response steps for the most common alerts. |
| 5. Hunt for the noise | Look for rapid, repeated exploit attempts and webshell activity on web servers. Automated attackers rarely slow down. |
The Fine Print
On November 17, 2026, Google Threat Intelligence Group reported that attackers are now using AI frameworks to discover and exploit vulnerabilities without a human in the loop, including a system called exploit_agent that turns new security patches into live attacks. The mechanics of intrusion have not changed, but their speed has, and a monthly patch cycle is no longer fast enough for internet-facing systems.
Frequently Asked Questions
My Take
AI did not invent a new type of attack. It removed the slowest part of an old one: the human. Patch diffing, exploit writing and scanning used to take skilled people days. Now they take an agent hours.
As a network security engineer, I read this report as a deadline, not a headline. The basics still win: patch what faces the internet, protect credentials, watch the logs. But they now have to run at the attacker's speed, which means automation on the defender's side too.
How fast can your team patch a critical internet-facing flaw today?
Sources:
- Help Net Security: Google finds AI agent writing and launching exploits with no human in the loop
- SecurityWeek: AI Framework Autonomously Weaponized Flaws in Real Attacks, Google Says
- The Hacker News: Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches
This article is published for information purposes only, to help readers understand AI-driven threats and how to defend against them.
Related Articles: