Google has disclosed that attackers attempted to replicate its artificial intelligence chatbot, Gemini, using more than 100,000 prompts in a single campaign.
In a report published on February 12, 2026, the company said the activity involved what it described as “distillation attacks”, in other words, repeated and structured questioning designed to extract the model’s reasoning patterns and internal logic.
Google said the goal of these campaigns was to replicate Gemini’s capabilities in competing systems. They described the activity as intellectual property theft.
A Google spokesperson told NBC News that the attackers appear to be “commercially motivated” actors, including private companies and researchers seeking a competitive advantage. The spokesperson added that the activity originated from multiple countries but declined to name specific suspects.
John Hultquist, Chief Analyst at Google’s Threat Intelligence Group (GTIG), said the scale of the attacks suggests similar incidents may soon affect smaller companies.
“We’re going to be the canary in the coal mine for far more incidents,” Hultquist said.
According to the report, attackers used authorised API access to send repeated queries to Gemini in an effort to reproduce how the system reasons across tasks, including in non-English languages.
Google said it identified one campaign that exceeded 100,000 prompts before taking corrective action.
In a separate disclosure, Google said state-linked threat actors have used Gemini to support various stages of cyberattacks.
Groups linked to China (APT31, Temp.HEX), Iran (APT42), North Korea (UNC2970) and Russia were identified in the report.
Google said these actors used Gemini for:
- Target profiling
- Open-source intelligence gathering
- Drafting phishing messages
- Translating text
- Debugging code
- Testing vulnerabilities
- Developing command-and-control systems
In one example cited by Google, a China-based actor fabricated a cybersecurity scenario and directed the model to analyse Remote Code Execution (RCE), Web Application Firewall (WAF) bypass techniques, and SQL injection test results against specific U.S.-based targets.
Google said there was no evidence that Gemini created new malware, but that it was used to accelerate existing attack processes.
Google stated that the model extraction campaigns do not represent a direct threat to user data.
The company said there is no indication that private conversations were exposed or that user information was stolen through these attacks.
However, cybersecurity analysts warn that AI tools can make phishing emails more convincing and malware development more efficient. For Nigerian internet users — particularly businesses, banks and government agencies that frequ
Nigeria has seen repeated cases of online fraud and phishing. The integration of AI tools into cybercrime could increase the scale and speed of such operations.
However, Google said it has taken several measures to address the issue, including disabling accounts linked to documented abuse, blocking related infrastructure, adjusting Gemini’s systems to better detect extraction attempts, and strengthening internal classifiers to make abuse detection more effective.
The company said it designs its AI systems with “robust security measures and strong safety guardrails” and regularly tests them to improve protection.
Google also acknowledged that large language models are inherently exposed because they are accessible online and respond to public queries.
The company described model extraction as “scalable” and said it poses a risk to the AI industry’s business model if left unchecked.
