Google reveals hackers use AI agents to automate cyberattacks
New research shows cybercriminals are shifting from using AI chatbots for advice to deploying autonomous agents that can scan, breach, and move through networks at machine speed.
The short version
- Attackers are using multi-agent frameworks like ARTEX and CAIRN to automate Active Directory attacks in seconds.
- Google found a framework called 'exploit_agent' that can autonomously find bugs, write exploits, and run them.
- Zero-day exploits jumped from 8 per month in 2025 to 22 in August 2026.
- AI is being used to create 'aged' fake social media accounts that mimic human scrolling habits for influence campaigns.
- Hackers are using AI-made images and fake recruiter profiles to target cryptocurrency firms.
Why it matters
This shift means security teams have significantly less time to react to breaches because attacks happen at machine speed. It also means that even simple tasks, like finding software bugs, can now be fully automated by malicious actors.
The shift to autonomous agents
Cybercriminals are moving beyond simply asking AI chatbots for help with their work. According to a report from the Google Threat Intelligence Group (GTIG) and Mandiant, hackers are now deploying AI agents that can perform complex tasks with very little human input. These agents can scan networks, break into systems, and move through internal infrastructures autonomously.
The research indicates that AI misuse has reached a "critical inflection point" across several sectors, including cybercrime, spying, and influence campaigns. By using these agents, threat actors can shorten their attack timelines and manage multi-stage intrusions at machine speed. This allows them to scale their operations much more effectively than manual methods allowed.
Google identified several open-source, multi-agent hacking frameworks being used in the wild. These include tools like ARTEX, CAIRN, CyberStrikeAI, and PentAGI. For example, CAIRN was used in intrusions against US aerospace targets, while a custom platform built on PentAGI scanned numerous organizations in Asia between late August and early October 2026.
Automating the exploit lifecycle
One of the most significant findings from the report is the emergence of tools that find and exploit software flaws without any human in the loop. Google discovered an exposed staging server hosting a framework called "exploit_agent." This system is designed to compare web software code before and after patches are applied to find bugs, write exploits, and then run those exploits on its own.
The report notes that these automated tools often trade stealth for speed. Because they move so quickly, they can leave behind noisy traffic, known names, and clear traces in the code. However, the research suggests that as these tools evolve, Google expects attackers to achieve better stealth while maintaining their high speed.
Real-world examples of this acceleration are already appearing. After two PaperCut flaws were exploited as zero-days in August 2026, an AI-assisted campaign hit hundreds of organizations. The attack reached domain admin status within minutes, illustrating how AI can compress the time between a vulnerability being discovered and it being fully exploited.
Financial and influence risks
Beyond direct network intrusions, AI is being used to facilitate large-scale financial theft and influence operations. Money-driven groups are stealing AI keys and cloud access to resell them through "LLM Mixers." Google's research into LLM Hijacking found fraudulent costs exceeding $100,000 in a single day, with some targets seeing their bills hit $1 million.
Influence groups are also integrating AI into their standard workflows to manipulate public opinion. One contractor used the Gemini API to build a platform capable of running over 100,000 social media accounts using cloud phone farms. These fake profiles were matured by AI that simulated typical human browsing and viewing behaviors before the accounts began posting political messages.
State-sponsored actors are also using AI to improve their phishing and infiltration tactics. A suspected North Korean cluster, MIDNIGHT NEPTUNE, used AI-made images and fake recruiter profiles to target cryptocurrency firms. They also used AI to hide their tracks by naming GitHub tokens after Microsoft Copilot to blend in with legitimate activity.
Defending against machine speed
Because AI allows attackers to chain methods together much faster, defenders have less time to respond to incidents. Google SecOps has released behavior-based rules specifically designed to detect the ARTEX framework. The goal is to provide security operations centers with the context they need to stop incidents quickly before they can spread.
Mandiant has released a five-layer defense guide to help teams protect their infrastructure. This guide advises organizations to lock down admin systems, use short-lived credentials, and split their networks. It also recommends monitoring for sudden spikes in API and identity activity, which can be a sign of an automated agent moving through a system.
For organizations using AI services, the report suggests specific proactive measures. Teams should set budget alerts to catch unexpected costs, audit their code for exposed keys, and treat any unplanned activation of an AI service as a high-priority warning. The underlying methods of hacking have not changed, but AI has significantly increased the velocity at which those methods can be executed.
What to watch
- How quickly AI agents can learn to bypass behavior-based detection rules.
- The growth of 'LLM Mixers' for reselling stolen cloud credentials.
Try it
Check your cloud service provider's dashboard to set up budget alerts for your AI services to catch unauthorized usage quickly.
Sources and links
First reported by CyberSecurityNews. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.