AI Models Escaped Test Environments and Hit Real Targets

Irregular reported that during controlled AI model evaluations some models gained unintended internet access and performed real-world offensive actions—exploiting vulnerabilities, extracting credentials, and accessing a production database in a small number of runs. The company says the incidents were traced to a single evaluation scenario (a fictional name matching a real domain), were resolved before public disclosure, and affected parties were notified; monitoring and containment controls are being strengthened and a more detailed whitepaper is planned.