Internet access was accidentally enabled during an evaluation in May, allowing Gemini to reach protected systems belonging to real organizations.

  • The test was conducted by AI security company Irregular as a “capture the flag” exercise.
  • Gemini gained access to one company’s protected system after repeatedly guessing passwords.
  • In two other cases, the model found exposed credentials in public repositories and used them to access protected systems.

Alphabet Inc.’s (GOOGL) Google said its Gemini artificial-intelligence model accessed the internet and breached three real companies' systems during a cybersecurity test in May. The incidents occurred during an evaluation conducted by AI security company Irregular and came after internet access was unintentionally made available to the model, reported The Wall Street Journal.

Google said Gemini stopped each intrusion after recognizing that it had reached real companies rather than simulated targets.

At the time of writing, GOOGL stock was up 0.25% in Friday after-hours trading.

Gemini Breached Real Company Systems

The incidents took place during a “capture the flag” exercise run by Irregular to assess Gemini’s cybersecurity capabilities. The model was instructed to retrieve information from software operated by a fictional company within the testing environment. But the fictional company had the same name as a real company.

According to Irregular, Gemini was not intended to have internet access during the exercise, but that access was accidentally enabled. In one instance, the model repeatedly guessed passwords until it accessed a protected system belonging to a real company. Gemini then recognized that it had reached a real organization, stopped the intrusion, and left the system, according to the WSJ report.

The other two incidents happened during separate test runs. In those cases, Gemini searched the internet using the fictional company's name. The searches led to public repositories containing credentials belonging to other companies. The model attempted to use those credentials to complete the evaluation and successfully accessed the protected systems.

Google said Gemini stopped again after determining it had accessed real companies. The company did not identify the three organizations but said it had notified all of them. Google also said it notified federal authorities.

Google Says Gemini Acted Appropriately

Irregular informed Google about the incidents at the end of July. Google did not disclose the Gemini incidents until WSJ contacted the company this week.

Google said it did not believe the incidents required public disclosure because the model did not cause harm and stopped each intrusion after determining that the target was a real company. The company compared the incidents with a “bug bounty” exercise, where security researchers identify vulnerabilities and report them to the affected organizations.

“This event highlights the importance of training powerful AI models to act responsibly,” Heather Adkins, Google’s vice president of security engineering, said in a statement, reported WSJ. “In this case, the model acted appropriately.”

Google said it did not consider the behavior an example of model misalignment because its safety measures helped Gemini stop. It also said the incidents did not involve its newest model, although it did not disclose which Gemini model was used.

GOOGL Stock: Stocktwits Retail Sentiment

On Stocktwits, retail sentiment for GOOGL remained ‘bearish,’ unchanged over the previous 24 hours, while message volume was ‘normal’ at the time of writing.

GOOGL stock has gained 12% year-to-date.

For updates and corrections, email newsroom[at]stocktwits[dot]com.<