Google Reports First Fully Automated AI Cyberattack as Hackers Use AI Agents to Exploit Software Flaws

Google has identified what its security researchers believe is the first cyberattack in which artificial intelligence handled the entire intrusion process without human intervention. The discovery highlights how quickly AI-powered hacking tools are changing the cybersecurity landscape.

According to a report published by Google Threat Intelligence Group and its Mandiant incident response unit on November 16, 2026, investigators uncovered a sophisticated framework that could identify software vulnerabilities, develop exploits and execute attacks automatically.

The operation involved a system called “exploit_agent,” which connected multiple AI agents with advanced and locally hosted language models. Researchers believe the activity was operated by Russian-speaking threat actors, although Google has not publicly identified a specific hacking group or government.

The investigation also found evidence that the attackers were interested in French websites associated with media, entertainment and education.

Google Discovers AI-Powered Cyberattack Operating Without Human Intervention

The most significant finding is the level of independence demonstrated by the attack framework. Earlier AI-assisted hacking operations generally involved humans directing the process while AI handled individual tasks.

This operation was different. Google’s investigators found evidence that the framework could move through vulnerability research and exploitation without requiring a person to approve each step.

The system used multiple AI agents to examine software changes, identify potential weaknesses and generate exploits. Once the framework identified an opportunity, it could attempt to execute the exploit automatically.

Google described the development as an important turning point in the use of artificial intelligence by cybercriminals.

The discovery does not mean that AI has independently launched attacks against every type of computer system. However, it demonstrates that attackers can now automate several technically demanding stages of an intrusion.

How the Exploit_Agent AI Framework Worked

The framework identified by Google researchers was named “exploit_agent.” Its primary function was to examine software vulnerabilities and turn available technical information into potential attack opportunities.

One of its central capabilities involved comparing software code before and after developers released security patches.

When a software vendor publishes an update, the changes can reveal what vulnerability the update was designed to fix. Security researchers use this process to understand weaknesses and protect systems.

Attackers can use the same information for malicious purposes. By examining the differences between patched and unpatched versions, the framework attempted to identify the underlying weakness and develop an exploit.

Also Read:  Mental health apps will clear your confusion now, its business is growing rapidly in India.

The AI system could then attempt to execute that exploit without waiting for a human operator to manually complete the process.

Google said the framework combined multiple AI agents with frontier and locally hosted language models. This structure allowed different components to handle separate tasks within the broader attack operation.

Hackers Targeted French Media, Entertainment and Education Websites

Investigators found information on a staging server suggesting that the operators were interested in French websites across three major sectors: media, entertainment and education.

The findings are important because these organizations frequently manage large amounts of sensitive information, digital accounts and internet-facing services.

However, Google did not identify the individual French organizations involved or publicly confirm the full extent of the information accessed.

Researchers also discovered Russian-language comments inside the operation’s configuration files. One configuration reference discussed API keys for automated exploitation.

These technical clues helped investigators assess the likely background of the operators. Google stopped short of attributing the activity to a named Russian hacking group or state agency.

The distinction matters because language found in malware or configuration files does not independently establish the identity or nationality of an attacker.

AI Framework Could Perform Reconnaissance and Install Webshells

The investigation uncovered additional scripts designed to support activity after an initial system intrusion.

These capabilities included internal reconnaissance, password guessing and installing webshells on compromised systems.

A webshell is a malicious script that can give an attacker remote access to a compromised web server. Once installed, it may allow unauthorized commands and further activity inside an organization’s network.

Researchers also found a particularly unusual capability: one script could remove webshells associated with rival attackers.

This suggests that the operators were not simply attempting to gain initial access. Their tools were also designed to manage access and maintain control after compromising a target.

The combination of automated vulnerability discovery, password attacks and post-intrusion activity makes the framework especially concerning for organizations that depend on internet-facing software.

Google Reports 22 Zero-Day Exploits in August 2026

The latest findings arrive as security researchers track an increase in exploited software vulnerabilities.

Google reported that it observed 22 exploited zero-day vulnerabilities in August 2026 alone. A zero-day vulnerability is a software weakness that is exploited before the affected vendor has released a fix.

Also Read:  Oppo users have fun! New AI features will be available in ColorOS 17, making photo editing and travel planning easier.

The figure compares with an average of approximately eight exploited zero-days per month in 2025 and around 11 per month earlier in 2026.

Another example involved two PHP vulnerabilities that vendors patched on October 21. Within days, hundreds of purported exploits appeared online.

Most of those exploits were reportedly nonfunctional and were believed to have been generated using AI.

Although unsuccessful exploit attempts do not necessarily lead to a security breach, the volume demonstrates how quickly automated tools can produce and distribute potential attack code after a vulnerability becomes public.

For security teams, this creates additional pressure to identify affected systems and apply patches before attackers can take advantage of the weaknesses.

Open-Source AI Hacking Tools Raise Additional Security Concerns

Google also identified activity involving open-source offensive security frameworks, including CAIRN and PentAGI.

Researchers observed these tools in activity targeting an aerospace organization in the United States and organizations across Asia.

The investigation also identified a China-linked threat actor developing an automated attack framework that could potentially support cyber espionage operations.

These findings indicate that the use of AI in offensive cybersecurity is not limited to one group or one type of operation.

Different threat actors are experimenting with automation to reduce the time required for reconnaissance, vulnerability research and credential collection.

Open-source frameworks can make advanced technical capabilities more accessible. However, their availability does not automatically mean every user is involved in criminal activity, as many security professionals use similar tools for authorized testing.

Why AI-Driven Cyberattacks Are Becoming More Difficult to Stop

Traditional cyberattacks often involve several stages that require human operators to make decisions. Security teams may detect unusual activity when attackers move between those stages or repeatedly interact with compromised systems.

Autonomous AI agents can reduce the time between those activities. They can analyze information, choose subsequent actions and troubleshoot certain problems without waiting for an operator.

That speed can make a significant difference when organizations have not yet installed security updates.

A vulnerability that becomes publicly understood today could attract automated exploitation attempts much sooner than security teams previously expected.

However, automation also creates opportunities for defenders. AI systems can help analyze security alerts, identify suspicious behavior and prioritize vulnerabilities that require immediate attention.

Also Read:  IRCTC Tatkal Tickets: You can get confirmed train tickets even on Diwali-Chhattis, do these small preparations before booking Tatkal.

The challenge is ensuring that defensive systems can respond at a speed that matches increasingly automated attacks.

What Businesses Should Do to Protect Against AI-Powered Cyberattacks

Organizations do not need to wait for a major security incident before reviewing their existing defenses. Google’s findings reinforce the importance of maintaining basic cybersecurity controls and reducing the time between vulnerability disclosure and patch installation.

Businesses should prioritize several security measures:

  • Apply security patches quickly: Internet-facing applications, servers and software should receive critical updates as soon as practical.
  • Monitor unusual network activity: Security teams should investigate unexpected connections, unusual login patterns and suspicious server behavior.
  • Strengthen access controls: Multi-factor authentication and least-privilege access can reduce the impact of stolen credentials.
  • Protect API credentials: API keys should be stored securely, rotated when necessary and monitored for unauthorized use.
  • Inspect web servers: Organizations should regularly check for unauthorized files, unexpected webshells and suspicious changes.
  • Maintain tested backups: Secure, offline or otherwise protected backups can help organizations recover from destructive incidents.
  • Use continuous security monitoring: Automated detection and incident response can help reduce the time between an intrusion and its discovery.

For companies operating in media, education, financial services and technology, these precautions are particularly important because their systems often depend on interconnected digital platforms.

AI Cybersecurity Debate Intensifies After Google’s Latest Findings

The discovery comes amid growing debate about how advanced AI systems could affect both offensive and defensive cybersecurity.

Anthropic introduced its Mythos cybersecurity-focused model in April 2026, while concerns have also emerged around autonomous AI agents accessing online platforms and development environments.

A separate incident involving an OpenAI agent and the development platform Hugging Face during the summer added to discussions about the risks of giving AI systems access to computer environments.

These developments have increased pressure on AI companies to strengthen safeguards while allowing legitimate security researchers to use advanced models for vulnerability discovery and defensive testing.

Google’s latest investigation adds another important example of how attackers are experimenting with increasingly autonomous systems.

For businesses and security professionals, the immediate concern is not simply that AI can write malicious code. It is that multiple AI agents can coordinate different stages of an intrusion, potentially reducing the time available to detect and stop an attack.

Enjoy reading Speaks Media?

Follow us on Google for the latest news.

Leave a Comment