Cybercriminals are beginning to use artificial intelligence to discover software vulnerabilities, develop exploits and execute cyberattacks without human intervention, according to a new warning from Google.
Google’s Threat Intelligence Group (GTIG) says it has identified the first instance of threat actors using large language model (LLM) frameworks to fully automate vulnerability discovery and exploitation.
In a report released by Google titled The era of autonomous attacks has arrived.” The security team described an AI-powered framework called exploit_agent, which can continuously examine software updates, identify vulnerabilities revealed by security patches, develop exploits and execute them automatically.
The discovery marks a significant shift in how cybercriminals can use AI. Rather than relying on human operators to investigate weaknesses and develop attack methods, attackers can now automate several stages of the process.
Google captured the threat with a stark warning: “The era of autonomous attacks has arrived.”
How autonomous AI attacks work
According to GTIG, the framework was discovered on an exposed staging server and was designed to work with multiple advanced and locally hosted AI models.
It uses a process known as differential analysis, comparing software code before and after a security update to identify what developers changed and which vulnerability they may have fixed.
Once it identifies a potential weakness, the system analyses the findings, develops an exploit and attempts to execute it without requiring a human to approve each step.
Researchers also discovered scripts on the server designed to support activities after a successful compromise, including internal reconnaissance, brute-force attacks, installing web shells and removing other attackers’ web shells.
The server’s activity suggested that its operator was interested in French websites connected to the media, entertainment and education sectors.
Google did not publicly identify the individual or group behind the framework.
The growing threat of AI-powered hacking
The discovery is part of a broader pattern in which cybercriminals are moving beyond using AI to write phishing messages or generate malicious code.
Google says threat actors are increasingly adopting multi-agent frameworks that can automate different stages of cyber operations, from scanning targets and identifying weaknesses to developing attack payloads.
The report identifies several frameworks being used or tested by threat actors, including ARTEX, CAIRN, CyberStrikeAI and PentAGI.
Some of these tools can coordinate specialised AI agents, generate attack payloads and adjust their tactics based on the results of previous actions.
Google also reported that the use of AI is helping attackers exploit vulnerabilities faster after software developers release security patches.
Between January and August 2026, the number of disclosed vulnerabilities and exploits nearly doubled, according to the report. Google also recorded a rise in zero-day exploits, which increased from an average of eight per month in 2025 to 11 per month in 2026, before reaching 22 in August.
The report warns that this growing speed is reducing the time organisations have to install security updates before attackers attempt to exploit known weaknesses.
Google says defenders must respond at machine speed
Google argues that traditional security approaches may struggle to keep pace with attacks that can operate continuously and execute multiple tasks automatically.
The company says organisations must increasingly adopt automated security systems capable of identifying vulnerabilities, prioritising risks and deploying fixes before attackers can exploit them.
To address this challenge, Google introduced Google AI Threat Defense, which combines Gemini and other advanced AI models with security capabilities from Wiz, automated vulnerability patching through CodeMender and threat intelligence from Mandiant.
However, Google also noted that current autonomous attack frameworks have weaknesses. Their speed and unrestricted execution can generate noticeable network activity and other forensic traces that make them easier for security teams to detect.
For businesses, financial institutions and cryptocurrency platforms, the warning reinforces the importance of timely software updates, strong access controls and continuous security monitoring. The threat is no longer limited to hackers using AI as an assistant. Attackers are beginning to build systems that can perform important parts of the hacking process on their own.
Read also: Italy fines Meta $887M over Facebook users’ photo access

