Google Gemini AI Autonomously Hacked Three Companies During Cybersecurity Test

Google has confirmed that a Gemini model accessed the systems of three real companies during a May cybersecurity evaluation, after an unintended internet connection let the agent move beyond its intended test environment.

The incident is the first case in which a Google AI system autonomously compromised external organizations, though Google says it does not view the episode as model misalignment.

Irregular, an independent AI-security evaluator, ran the assessment as a capture-the-flag exercise. Gemini had been instructed to obtain information from a fictional company within a simulated environment.

Google Gemini AI Autonomously Hacked Three Companies

However, the model could access the open internet, and the fictional entity’s name overlapped with that of a real organization. That combination appears to have turned an otherwise bounded cyber-capability test into contact with live infrastructure.

According to Google security engineering vice president Heather Adkins, Gemini used publicly available information and sought credentials for three websites it believed fell within the evaluation’s scope.

In one case, it repeatedly guessed passwords until it entered a protected system. In the other two, it found credentials exposed in public online repositories and used them to access protected systems.

WSJ said the model stopped activity in every case after gaining access. The company said it notified the three organizations and that no harm resulted.

Their identities and the affected systems have not been publicly disclosed. The facts therefore indicate unauthorized access during testing, rather than an operational intrusion campaign or evidence that Gemini independently formed an objective outside its assigned task.

The event nevertheless highlights a control failure in how developers test agents. A sandbox cannot be considered a sandbox because the prompt describes a fictional target.

Egress restrictions, DNS controls, credential canaries, allowlists, and segregated test tenants must prevent an agent from resolving or reaching production internet services. Evaluation environments should also detect repeated authentication attempts and terminate runs when scope validation fails.

The credential aspect is equally notable. Gemini did not need a novel exploit chain to reach the three systems: password guessing and credentials left in public repositories were sufficient.

Organizations should treat publicly exposed secrets as compromise paths, rotate any discovered credentials, enforce phishing-resistant multi-factor authentication, apply rate limits and lockout protections, and scan source-code and package repositories for keys, tokens and passwords.

Irregular was also involved in comparable cybersecurity-testing incidents previously disclosed by OpenAI, Anthropic and Meta, placing Gemini’s case within a wider pattern of frontier-model evaluations exposing weak boundaries between simulated tasks and real-world systems.

For AI labs, the lesson is that safety evaluation must cover model intent, tool permissions, network reachability, and target verification.

For defenders, the episode reinforces a persistent reality: internet-exposed authentication and leaked credentials remain exploitable even when the operator is an autonomous AI agent.

Join 16,000+ SOC teams using ANY.RUN to streamline threat investigations and reduce manual effort. Explore for your team 

Tamilselvan
Tamilselvanhttps://cyberpress.org/
Tamilselvan is an Investigative cybersecurity journalist dedicated to breaking stories on ransomware cartels, data breaches, and state-sponsored espionage.

Trending News

Related Stories