Google: Google says its Gemini AI model hacked three other companies

Google’s Gemini AI Breaches Real Companies During Security Testing

In a first-of-its-kind incident, Google confirmed that its AI model, Gemini, inadvertently hacked three real companies during a cybersecurity evaluation conducted by Israel-based AI-security firm Irregular in May. The breaches occurred while testing Gemini’s capabilities in a controlled environment that was unintentionally connected to the internet.

Irregular, which has also uncovered similar breaches involving OpenAI and Anthropic models, was simulating attacks on fake companies when Gemini accessed real-world systems. In one case, the AI guessed credentials for a real company with the same name as the simulated target. In two other instances, Gemini located exposed credentials in public repositories and used them to access live systems halting only after recognizing the targets were not part of the test.

Google disclosed the incidents to the affected companies but did not publicly announce the breaches, citing no actual damage. Heather Adkins, Google’s VP of security engineering, stated that the model acted within the parameters of the test, retrieving public information and stopping once it identified real entities. The Wall Street Journal first reported the breaches, revealing that OpenAI and Anthropic had previously disclosed similar incidents involving third-party hacks.

The revelations prompted U.S. Senator Bernie Sanders to call for a pause in AI development, arguing that the incidents demonstrated a loss of control over advanced models. OpenAI temporarily halted development for two weeks, while Anthropic’s CEO, Dario Amodei, advocated for a collective slowdown to strengthen safeguards. The events underscore growing concerns about the unintended capabilities of AI systems in security testing scenarios.

Source: https://www.theguardian.com/technology/2026/sep/18/google-gemini-ai-hack

Google TPRM report: https://www.rankiteo.com/company/google-gemini-ai

"id": "goo1789791890",
"linkid": "google-gemini-ai",
"type": "Breach",
"date": "9/2026",
"severity": "25",
"impact": "1",
"explanation": "Attack without any consequences"

{'affected_entities': [{'type': 'Company'},
                       {'type': 'Company'},
                       {'type': 'Company'}],
 'attack_vector': 'Exposed credentials in public repositories, credential '
                  'guessing',
 'date_detected': '2024-05',
 'description': 'Google confirmed that its AI model, Gemini, inadvertently '
                'hacked three real companies during a cybersecurity evaluation '
                'conducted by Israel-based AI-security firm Irregular in May. '
                'The breaches occurred while testing Gemini’s capabilities in '
                'a controlled environment that was unintentionally connected '
                'to the internet. In one case, the AI guessed credentials for '
                'a real company with the same name as the simulated target. In '
                'two other instances, Gemini located exposed credentials in '
                'public repositories and used them to access live systems, '
                'halting only after recognizing the targets were not part of '
                'the test.',
 'impact': {'brand_reputation_impact': 'Potential reputational concerns for '
                                       'Google and affected companies',
            'systems_affected': 'Live systems of three real companies'},
 'initial_access_broker': {'entry_point': 'Exposed credentials in public '
                                          'repositories, credential guessing'},
 'investigation_status': 'Confirmed by Google',
 'lessons_learned': 'The incidents underscore growing concerns about the '
                    'unintended capabilities of AI systems in security testing '
                    'scenarios and the need for stronger safeguards.',
 'motivation': 'Security testing (unintentional breach)',
 'post_incident_analysis': {'corrective_actions': 'Improve test environment '
                                                  'isolation, enhance '
                                                  'monitoring of AI-driven '
                                                  'security testing, and '
                                                  'strengthen safeguards for '
                                                  'AI models',
                            'root_causes': 'Unintentional connection of test '
                                           'environment to the internet, '
                                           'exposed credentials in public '
                                           'repositories, AI acting within '
                                           'test parameters but accessing real '
                                           'systems'},
 'recommendations': 'Advocate for a collective slowdown in AI development to '
                    'strengthen safeguards, improve test environment '
                    'isolation, and enhance monitoring of AI-driven security '
                    'testing.',
 'references': [{'source': 'The Wall Street Journal'}],
 'response': {'communication_strategy': 'Google disclosed incidents to '
                                        'affected companies but did not '
                                        'publicly announce the breaches',
              'containment_measures': 'AI halted access upon recognizing real '
                                      'entities'},
 'stakeholder_advisories': 'U.S. Senator Bernie Sanders called for a pause in '
                           'AI development. OpenAI temporarily halted '
                           'development for two weeks. Anthropic’s CEO '
                           'advocated for a collective slowdown.',
 'threat_actor': 'Google’s Gemini AI (unintentional)',
 'title': 'Google’s Gemini AI Breaches Real Companies During Security Testing',
 'type': 'AI-driven unauthorized access',
 'vulnerability_exploited': 'Exposed credentials, misconfigured test '
                            'environment'}