Security
Google discovers autonomous AI hacking tool with Russian code comments
Security researchers found software that automatically identifies vulnerabilities, writes exploits, and executes attacks without human intervention.
Key takeaways
- Google discovered the first known AI hacking tool that autonomously finds vulnerabilities, writes exploits, and executes attacks without human involvement
- The tool, named exploit_agent, monitors software patches, compares code versions to identify flaws, then generates and runs attacks automatically
- Russian-language comments were found in the tool's files, discovered on an exposed staging server along with data from French media and education targets
- North Korean hackers are using AI-generated recruiter profiles and disguising backdoors as legitimate AI coding assistants to target cryptocurrency firms
- Zero-day exploits jumped from an average of 11 per month to 22 in August 2026, while overall vulnerability disclosures nearly doubled in the first eight months of the year
Autonomous exploitation without human oversight
Google's Threat Intelligence Group has identified what it describes as the first fully autonomous AI-powered hacking tool capable of discovering and exploiting software vulnerabilities without human intervention, according to the company's third-quarter threat report.
The tool, which Google named "exploit_agent," was discovered on an exposed staging server accessible via the internet. Security researchers found the software integrated with multiple large language models, including both commercial platforms and locally-run AI systems.
The autonomous workflow begins when the tool monitors popular web software for newly released security patches. Upon detecting an update, it performs a code comparison between the patched and unpatched versions to identify the specific vulnerability that was fixed. The system then generates exploit code targeting that weakness and executes the attack automatically.
Google's report noted that comments within the tool's files were written in Russian. When translated, one comment read: "LLM API keys for auto-exploitation (exploit_agent pilot)." Another instructed users to "Enter the keys yourself — right here, not in the chat. Both lines are required."
Why it matters
This discovery represents a significant escalation in offensive AI capabilities. While security researchers have long used automated tools to assist with vulnerability discovery, a system that completes the entire attack chain—from patch analysis to exploitation—without human guidance crosses a threshold that security experts have warned about. The finding validates concerns raised by regulators about autonomous AI systems threatening critical infrastructure, from banking systems to software supply chains.
Broader attack infrastructure
The staging server contained additional attack tools beyond the autonomous exploit generator. Google identified scripts for internal network reconnaissance, password brute-forcing capabilities, and webshell deployment mechanisms that would allow persistent access to compromised systems.
Evidence on the server suggested the operator had targeted French websites in media, entertainment, and education sectors. The report documented one instance where the tool analyzed access control vulnerabilities in Spectra, a WordPress plugin, though Google did not confirm whether this led to a successful compromise.
Google's data shows vulnerability exploitation activity nearly doubled between January and August 2026. Zero-day exploits—attacks targeting previously unknown flaws—averaged eight per month in 2025 and eleven per month through most of 2026, before spiking to 22 in August.
Cryptocurrency sector targeted with AI-enhanced attacks
The report also detailed North Korean threat actors using AI-generated content in social engineering campaigns against cryptocurrency companies in North America and Europe. The group, tracked as MIDNIGHT NEPTUNE, created recruiter profiles on professional networking sites using AI-generated images marked with Google's SynthID watermarks.
Once attackers gained access to target networks, they modified configuration files in internal code repositories. These changes caused a backdoor called SOMBERMEME to download when developers opened Claude Code, an AI coding assistant. The attackers also created GitHub access tokens named after Microsoft Copilot to disguise malicious activity as legitimate AI tool usage.
According to blockchain forensics firm TRM Labs, North Korea-linked entities have stolen approximately $12.8 billion in cryptocurrency to date. A recent Binance breach resulted in roughly $6.3 billion in losses attributed to North Korean hackers by the FBI and forensics firms.
These findings were first reported by CoinDesk, based on Google's Threat Intelligence Group quarterly report released Monday.