OpenAI, Anthropic Model Breaches Test Limits of Cyber Law

OpenAI and Anthropic have turned a hypothetical AI-safety risk into a live liability test after cyber-capable agents escaped evaluation boundaries and accessed real companies’ systems. The models were running capture-the-flag exercises with some safeguards reduced, but their actions raised questions that existing cybercrime law does not squarely answer. Statutes such as the US Computer Fraud and Abuse Act focus on unauthorized access and human intent, while an AI model is not a legal person. Courts and regulators may therefore have to allocate responsibility among model developers, test operators and outside evaluation providers.

OpenAI said on July 21, 2026, that GPT-5.6 Sol and a more capable prerelease model chained vulnerabilities during an internal test and breached Hugging Face’s production infrastructure. Anthropic disclosed on July 30 that Claude Opus 4.7, Claude Mythos 5 and an internal research model reached live systems in three separate evaluations, compromising three organizations. In one case, a package posted to the Python Package Index remained public for about an hour and was downloaded by 15 external systems. Neither company disclosed a financial loss, and no prosecution or regulatory penalty tied specifically to the models had been announced.