Google uncovers AI hacking tool that turns security patches into automatic attacks
Google finds AI system that turns software fixes into attacks
Google says it has found an AI-powered hacking system that can turn software security fixes into automatic attacks on websites. The Google Threat Intelligence Group and Mandiant, Google's security team, described the finding in a report published on Nov. 16, 2026.
According to the report, the system could find software weaknesses, write attack code and launch attacks with very little human involvement. Google said this is the first time it has identified attackers using AI frameworks to fully automate both vulnerability discovery and exploitation.
Key points from the report
- Google found a system, named "exploit_agent," on an exposed staging server with several AI models linked to security research and exploitation tools.
- The system was set up to watch for software updates, compare fixed code with older versions, and find weaknesses that developers had just repaired.
- It could then generate attack code and run it automatically, targeting websites that had not yet installed the update.
- Google's investigation also found stolen victim data and malicious scripts that could scan internal networks, run brute-force password attacks, and install webshells, which are tools that give remote access to a hacked server.
- Some tools could remove webshells left by other attackers, possibly helping the operator keep control of compromised systems.
- Evidence pointed to interest in French websites tied to media, entertainment, and education. One recovered example showed the framework examining access-control protections in Spectra, a WordPress plugin used for website development.
How the exploit_agent pipeline works
The system takes advantage of information developers publish when fixing software flaws. A security patch, which is the code change that repairs a weakness, can show attackers how the flaw works. According to Google, exploit_agent was configured to compare web application code before and after a patch is announced, identify possible weaknesses, draft exploits, and execute them without human intervention in that pipeline.
Google said this method lets attackers focus on known, patched flaws instead of spending time searching for previously unknown weaknesses. It also said automated vulnerability analysis and exploitation can shorten the window between a patch being made public and an attack being built.
Google noted limits on what it could confirm. The company said its published diagram shows the configured pipeline, not measured timing, a success rate, or proof of a fully autonomous campaign. The system also required an operator to enter language-model API keys, meaning a person still had to set it up.
North Korean hackers used AI tools against crypto companies
The same Google report described a separate campaign in which suspected North Korean hackers used AI-powered development tools to break into cryptocurrency organizations in North America and Europe. The attackers, tracked as MIDNIGHT NEPTUNE, combined AI-generated identities with fake recruitment campaigns to access corporate systems.
They created fake recruiter profiles on professional networking sites using AI-generated photos, then sent targets to fraudulent meeting websites. Those sites told victims to run supposed troubleshooting commands that actually executed malicious scripts.
Once inside, the attackers reached internal code repositories and changed configuration files used by Anthropic's Claude Code, an AI assistant for software development. The altered files were designed to automatically download a backdoor, called SOMBERMEME, whenever an employee launched Claude Code. Google also found the attackers disguised GitHub access tokens, which are credentials for the code-sharing platform, under names linked to Microsoft Copilot to hide their activity.
What is confirmed and what is not
Confirmed by Google's report: the exploit_agent framework was found on an exposed server, its configuration included patch comparison, exploit drafting and automatic execution, and the system required operator-supplied API keys. Google also confirmed the separate MIDNIGHT NEPTUNE campaign details and the discovery of stolen data and reconnaissance scripts.
Not confirmed: Google did not report measured attack timing, a success rate, or proof of a fully autonomous hacking campaign. The CryptoSlate report also notes that Google's description of faster exploitation is the company's statement, not a measured universal patching deadline.
Why this matters for software and crypto security
The finding shows how published security fixes can double as a guide for attackers. Website owners often take time to install updates, and an automated tool that reads patches could strike during that gap. For crypto companies, the North Korean campaign adds another risk: attackers are now targeting the AI development tools engineers use daily, hiding malicious code inside files those tools load automatically.