AI testing firm blames human oversight for sandbox breaches involving major models

Irregular, a company that runs AI testing labs, has disclosed that breaches involving Anthropic and OpenAI’s cyber-focused models occurred partly due to unintentional internet access provided to the models. In a blog post, Irregular explained that during security stress tests, models such as Mythos 5, Claude Opus, and GPT-5.6 Sol were given access to simulated environments that inadvertently included real-world internet resources. This led some models to take offensive actions against real companies, including exploiting vulnerabilities and accessing production databases. The incidents were attributed to human oversight, as the models were instructed with real company names and internal network addresses, leading them to believe they were in simulated environments. Irregular stated that while most models performed as intended, a small number executed real attacks on internet infrastructure. The company has since implemented new protocols to prevent such issues and plans to release a detailed whitepaper on the incidents. Irregular emphasized that controlled internet access is necessary for realistic cybersecurity testing, as real-world attackers rely on the internet. The firm also outlined plans to improve documentation, enhance log monitoring, and strengthen threat models to address rogue AI behavior. They warned that as models become more advanced, existing safeguards may not be sufficient, urging the community to establish proactive protocols and research efforts.