Google’s Threat Intelligence Group and Mandiant have released new research documenting what they describe as a significant escalation in AI-powered cyberattacks, including the first observed use of AI frameworks to fully automate vulnerability discovery and exploitation without human intervention.
Researchers identified an exposed server running a multi-agent framework called exploit_agent, designed to continuously analyze software patches, identify vulnerabilities, generate exploits and execute attacks autonomously. Evidence collected from the server indicated that the operator was particularly interested in targeting French media, entertainment and educational websites.
Google also documented attackers using autonomous offensive security frameworks, including CAIRN and PentAGI, during real-world intrusions.
The company warned that these tools can dramatically accelerate attacks by automating multiple stages of an intrusion, reducing the time defenders have to identify and address vulnerabilities.
John Strand, Owner, Black Hills Information Security
“We’ve been using agentic frameworks for exploit development, attack surface enumeration, and reconnaissance at Black Hills Information Security for about a year now. And I think one of the biggest changes we’ve seen over the past couple of months is that these models are becoming increasingly effective on much smaller hardware and memory footprints.
“We’re now getting to the point where systems with 126 GB of memory can run models that are very effective at conducting these types of attacks. Sure, they’re slower than the frontier models, but they absolutely work. And that changes things dramatically because you no longer need massive infrastructure to start automating significant portions of an attack.
“In fact, I was doing some research on Shodan today and identified what appeared to be roughly 2,000 AI penetration testing platforms exposed directly to the internet. That number is fascinating, and I expect it to continue growing as this technology becomes more widely adopted.
“But here’s the part that really caught my attention. A number of these systems appeared to be compromised computers that were running these agentic attack frameworks on top of them. Think about what that means. We’re potentially seeing attackers deploy autonomous offensive capabilities directly onto compromised infrastructure.
“We’re absolutely seeing an explosion in these capabilities. And I don’t think this is some temporary trend. I think agentic AI is going to become the new normal for offensive cybersecurity operations over the next five years, and defenders need to start preparing for that reality right now.”
Stop me if you have heard this before. If your defenses do not include AI, they need to or else bad things will happen.
Related
This entry was posted on November 17, 2026 at 3:00 pm and is filed under Commentary with tags Google. You can follow any responses to this entry through the RSS 2.0 feed.
You can leave a response, or trackback from your own site.
Google identifies fully automated AI cyberattacks operating without human intervention
Google’s Threat Intelligence Group and Mandiant have released new research documenting what they describe as a significant escalation in AI-powered cyberattacks, including the first observed use of AI frameworks to fully automate vulnerability discovery and exploitation without human intervention.
Researchers identified an exposed server running a multi-agent framework called exploit_agent, designed to continuously analyze software patches, identify vulnerabilities, generate exploits and execute attacks autonomously. Evidence collected from the server indicated that the operator was particularly interested in targeting French media, entertainment and educational websites.
Google also documented attackers using autonomous offensive security frameworks, including CAIRN and PentAGI, during real-world intrusions.
The company warned that these tools can dramatically accelerate attacks by automating multiple stages of an intrusion, reducing the time defenders have to identify and address vulnerabilities.
John Strand, Owner, Black Hills Information Security
“We’ve been using agentic frameworks for exploit development, attack surface enumeration, and reconnaissance at Black Hills Information Security for about a year now. And I think one of the biggest changes we’ve seen over the past couple of months is that these models are becoming increasingly effective on much smaller hardware and memory footprints.
“We’re now getting to the point where systems with 126 GB of memory can run models that are very effective at conducting these types of attacks. Sure, they’re slower than the frontier models, but they absolutely work. And that changes things dramatically because you no longer need massive infrastructure to start automating significant portions of an attack.
“In fact, I was doing some research on Shodan today and identified what appeared to be roughly 2,000 AI penetration testing platforms exposed directly to the internet. That number is fascinating, and I expect it to continue growing as this technology becomes more widely adopted.
“But here’s the part that really caught my attention. A number of these systems appeared to be compromised computers that were running these agentic attack frameworks on top of them. Think about what that means. We’re potentially seeing attackers deploy autonomous offensive capabilities directly onto compromised infrastructure.
“We’re absolutely seeing an explosion in these capabilities. And I don’t think this is some temporary trend. I think agentic AI is going to become the new normal for offensive cybersecurity operations over the next five years, and defenders need to start preparing for that reality right now.”
Stop me if you have heard this before. If your defenses do not include AI, they need to or else bad things will happen.
Share this:
Like this:
Related
This entry was posted on November 17, 2026 at 3:00 pm and is filed under Commentary with tags Google. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.