AI News: Gemini AI Hacked 3 Real Companies During Security Test

Principali approfondimenti:

• AI News: Gemini accessed three companies during a cybersecurity test.
• Gemini stopped after realizing that the targets it had accessed were real companies.
• Unintended internet access let Gemini reach protected external systems.

Google confirmed Gemini accessed three real companies during a May cybersecurity evaluation run by AI security firm Irregular. The AI News disclosure emerged Sept. 18, and Google said Gemini stopped after recognizing real targets.

Irregular stated that unintended internet access allowed the model to reach live systems outside the simulation.

The incident mattered because Google is expanding Gemini across consumer, enterprise, and cybersecurity products. Alphabet said in July that the Gemini app had 950 million monthly active users.

The company also revealed that nearly 90% of Fortune 100 companies used Gemini Enterprise. The episode, therefore, tested controls around an increasingly important Google AI platform.

AI News: Google Confirms Three Gemini Intrusions

As per Reuters’ report, Google Vice President of Security Engineering Heather Adkins confirmed the three incidents in a statement. She said Gemini found public information and guessed credentials for websites it considered part of the evaluation.

Adkins said the model stopped in all three cases after identifying real targets. Google also notified the affected companies and worked with Irregular on revised testing procedures. Google said the activity caused no harm to the companies.

Irregular said it identified interactions where unintended internet access enabled offensive actions against real systems. The company said it remediated the testing issues and contacted affected parties.

Google did not publish a standalone incident report before the media inquiries. Irregular said relevant laboratories received notifications in late July. The security company published its broader incident review on Aug. 14.

How Gemini Reached Systems Outside the Test

The evaluation used a capture-the-flag format that asked Gemini to retrieve information from simulated software. Irregular describes these tests as controlled challenges with defined objectives and isolated infrastructure.

The Wall Street Journal ha riportato that a fictional target shared a name with a real company. Internet access also remained available when the test environment should have blocked it.

Gemini guessed a password and gained access to a protected system, the Journal reported. In two other cases, it found credentials stored in public repositories. The model then used those credentials to access systems operated by real companies.

Irregular’s Aug. 14 review did not identify Google or Gemini by name. However, it acknowledged that unintended internet access had enabled some models to take offensive external actions. The company said it added safeguards to prevent similar incidents.

Google AI Safety Controls After the Test

The incident came as Google increased Gemini’s role in cybersecurity. Google introduced Gemini 3.5 Flash Cyber in July for vulnerability detection, validation, and remediation.

Google DeepMind’s Frontier Safety Framework also treats advanced cyber capability as a tracked risk domain. Its April 2026 update added earlier capability thresholds and expanded risk-management procedures.

Google said Gemini 3.5 Flash remained below its Cyber Critical Capability Level. The related model card said prior Gemini models had crossed an earlier cyber alert threshold.

Those assessments measure model capability under controlled conditions, not the exact circumstances behind the May incidents. Still, the disclosures show why containment matters alongside model-level safety testing.

Alphabet’s financial disclosures also show Gemini’s growing commercial role. Second-quarter Google Cloud revenue reached $24.8 billion, up 82% from one year earlier. Alphabet attributed the increase partly to enterprise artificial intelligence infrastructure and solutions.

Alphabet’s June filing separately warned about the exposure of confidential data tied to the use of artificial intelligence. The company said such exposure could bring regulatory investigations, enforcement actions, and higher compliance costs.

Its filing also linked those risks to potential harm to the business, reputation, finances, and operations. The filing did not mention Irregular. Google’s later statement provided the company’s public account of what occurred.

AI News: Irregular Says Testing Gaps Were Fixed

Irregular said it had remediated the issues that allowed models to reach live systems. It also said additional safeguards now prevent similar evaluation failures.

The firm’s broader research describes network boundaries, authentication, sandboxing, and controlled exposure as core evaluation controls. Its FrontierCyber benchmark uses real systems under instrumented and repeatable conditions.

Google has not identified which Gemini version participated in the May evaluation. The company has not named the three affected companies.

Alphabet’s next reporting quarter ends Sept. 30, 2026. Its subsequent Form 10-Q offers the next routine disclosure point for related risk updates. Investors can compare any revised language with Alphabet’s June filing.

L'articolo AI News: Gemini AI Hacked 3 Real Companies During Security Test è apparso per primo su The Coin Republic.