Google Confirms Cyberattacks Launched by AI Without Human Help in First Real-World Case

Google Confirms AI Cyberattacks Without Human Help in Real Case

Google has confirmed the first real-world case of hackers using an artificial intelligence system to turn software security updates into live cyberattacks without human help during the attack process.

The finding was published on November 16, 2026, by the Google Threat Intelligence Group and its cybersecurity unit, Mandiant. Just two months earlier, Google reported that while criminals were experimenting with AI tools, no fully automated attack system had yet been seen in real operations. According to Google, that line has now been crossed.

Researchers discovered the tool, named exploit_agent, after finding an exposed staging server used by a suspected Russian-speaking operator. Unlike earlier cases where hackers typed questions into a chatbot for coding help, exploit_agent was built as an automated pipeline connected to several commercial and locally run AI models.

Readers may wonder why AI companies do not simply block requests to write attack code. While commercial AI providers build safety filters into their products and actively shut down abusive accounts, automated tools like exploit_agent connect directly to AI models through developer APIs, often using stolen API keys, or run locally hosted open-weight models on private servers. Because code comparison and debugging look similar to legitimate security testing, these technical access methods can slip past restrictions designed for everyday public chatbots.

Once switched on, the system worked on its own with no human in the loop. It watched popular web software for security updates, compared the old and new code to see what changed, worked out what weakness the update fixed, wrote working attack code, and launched that attack automatically against websites online.

In its report, Google said it “identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time,” warning that this points to attacks happening at machine speed.

For website owners and businesses, the biggest concern is how exploit_agent turns routine software updates into an attack guide. Whenever a software company releases a security fix, the updated code shows what was broken. Security specialists call comparing the old and updated files “patch diffing”.

In the past, skilled human hackers needed days or weeks to study an update, understand the flaw, and build a tool to target sites that had not installed it yet. By handing that code comparison to an AI agent, attackers can shrink the window between a public security update and real-world attacks. Google noted that because the changes inside an update are small and clear, “threat actors may find it more efficient to use LLMs to rapidly weaponize patched n-day vulnerabilities” rather than hunting for brand-new secret flaws.

This shift also changes which websites are likely to be targeted. Human hackers typically spend their time researching high-value companies, which often meant smaller blogs, local businesses, and community sites were overlooked. Automated agents do not get tired or choose targets manually, so they scan the internet broadly and test any unpatched website running popular software.

Evidence on the exposed server showed the system was used in real attacks, not just lab tests. Google confirmed the server held data stolen from victim organizations, along with follow-up scripts built to scan internal networks, guess passwords, install remote-control webshells, and even delete webshells left behind by rival hackers. Data from the server showed the operator was targeting French websites across the media, entertainment, and education sectors.

Everyday website software was also inspected by the system. One technical example in Google’s report showed exploit_agent examining security checks inside version 2.19.25 of Spectra, a popular WordPress page-building plugin used on more than one million sites.

Even so, Google’s findings draw a clear line between automated attacks and human preparation. The AI did not plan or start the operation by itself. A human operator set up the server, left Russian-language notes telling the administrator to paste AI API keys directly into a settings file, and prepared the follow-up scripts.

Important details also remain unknown. Google did not name the exact AI models connected to the server, link the server to a known cybercrime gang, say how many organizations had data stolen, or confirm whether inspecting the Spectra plugin led to any hacked WordPress sites.

The discovery also does not mean every website or every software update is suddenly defenseless. Google stressed a major weakness in current autonomous attack tools: while they speed up attacks, they “currently trade stealth for velocity”.

Because automated agents scan and test websites rapidly and repeatedly, they create noisy, unusual internet traffic that security tools can spot and block far more easily than a careful human hacker. In related cases tracked by Google, AI-written attack scripts also frequently failed to work, creating loud network noise without breaking in.

Still, exploit_agent is part of a wider shift across cybercrime and state-backed hacking. Google reported that attackers are increasingly using open-source AI testing frameworks such as CAIRN, PentAGI, CyberStrikeAI, and HexStrike AI in real intrusions, including attacks aimed at the U.S. aerospace sector. Days earlier, the FBI and Cisco Talos warned that hackers were using another open-source AI framework, ARTEX, to take over networks from initial entry to administrator control in under three hours.

For website owners, businesses, and everyday users across the UK, the United States, and Europe, the takeaway is simple: waiting weeks to install routine updates leaves a dangerous opening when attackers can read patches at machine speed.

WordPress and small website owners should turn on automatic updates for trusted plugins and themes so security fixes install within hours rather than weeks. Unused plugins should be deleted completely, since deactivated plugins still sit on your server, and public websites should run behind a web application firewall that blocks rapid automated scanning.

Businesses and software teams should treat security updates for internet-facing systems as urgent and watch server logs for sudden bursts of automated probing right after major patches are released. Developers using AI coding tools should also check project settings files before opening shared code folders, as Google separately found a suspected North Korean group hiding malware triggers inside .claude/settings.json files to infect developers launching Claude Code.

For everyday internet users, keeping web browsers, smartphones, and home routers set to install security updates automatically remains the most reliable way to stay ahead of automated attacks.

More from Threat Intel

Community discussion

Join the conversation

0 comments

Be the first to join in

There are no comments yet. Sign in to start a thoughtful discussion.