Three Real Companies Were Breached by Gemini AI in Google’s May Security Test

The tech giant says the May exercise showed Gemini could autonomously breach live corporate systems

Google has confirmed that its Gemini artificial intelligence model breached the systems of three real companies during a security test in May. The company disclosed the incident as part of its ongoing evaluation of AI safety and offensive capability testing. The confirmation adds to mounting evidence that advanced AI models can conduct complex cyber operations without direct human guidance.

The test appears to have been designed to probe how far an AI system could go when tasked with penetrating live corporate networks. Rather than simulating attacks against sandboxed environments, Gemini reportedly interacted with actual company infrastructure. That distinction matters. Security researchers have long debated whether lab-based AI red-teaming translates into real-world risk.

Google has not detailed which three companies were targeted or what specific vulnerabilities Gemini exploited. The lack of public specifics leaves open questions about the scope of the breaches and whether any sensitive data was accessed or exposed. It is also unclear whether the companies involved consented to the test in advance or were selected without prior notification.

The episode fits into a broader industry conversation about dual-use AI models. The same reasoning and coding abilities that make large language models useful for defensive security work can also be turned toward offensive hacking. Google’s own safety teams have previously flagged this tension in public research notes on frontier model risk.

AI companies including Google, OpenAI, and Anthropic have expanded internal red-teaming programs over the past two years. These programs test whether their models can be misused for tasks like malware generation, phishing campaign design, or network intrusion. A successful autonomous hack against real companies represents a notable data point in that effort.

The timing of the disclosure, months after the actual May test, suggests Google may have used the interval to assess the implications internally. It is common practice for major AI labs to delay disclosure of security findings until mitigations are in place. Whether Google has since restricted Gemini’s capabilities in this area has not been confirmed.

For the wider technology and digital asset sectors, the news underscores a persistent worry. As AI systems grow more capable, the line between security research and genuine offensive risk narrows. Crypto exchanges, custodians, and blockchain infrastructure providers rely heavily on network security. Any evidence that AI models can autonomously breach corporate systems is directly relevant to how these firms assess future threats.

Market Impact

The disclosure is likely to intensify scrutiny of AI models used in cybersecurity contexts, including tools deployed by exchanges and custodians to monitor for intrusions. Crypto firms, which manage large pools of digital assets and rely on layered security architecture, may face renewed pressure to audit how AI-assisted tools interact with their systems. Investors in AI-linked equities and tokens tied to cybersecurity themes could see increased attention to safety disclosures from major AI developers going forward.

Regulators focused on AI governance may also cite the incident as evidence supporting stricter oversight of frontier model testing and deployment. For now, there is no indication that the test caused financial losses at the affected companies or that any crypto-specific systems were involved.

Google's confirmation that Gemini breached three real companies during testing marks a significant moment in the debate over AI-driven cyber risk. With key details still undisclosed, the full scope of the exercise and its consequences remain to be seen.

Frequently Asked Questions

What exactly did Google confirm about Gemini?

Google confirmed that its Gemini AI model successfully hacked three real companies during an internal security test conducted in May.

Which companies were affected?

Google has not publicly named the three companies involved in the test.

Was any data stolen or exposed during the test?

Specific details about data access or exposure during the incident have not been disclosed by Google.

Why does this matter for the crypto industry?

Crypto exchanges and custodians depend on strong network security, so evidence that AI can autonomously breach corporate systems is relevant to how these firms evaluate emerging threats.

Has Google restricted Gemini's capabilities as a result?

It has not been confirmed whether Google has placed new restrictions on Gemini following the test.