When AI Becomes The Threat: A Blueprint For Managing Frontier AI Risk

Kevin Lynch, CEO and board member at Optiv, a cyber advisory and solutions leader.

​Since the introduction of ChatGPT in November 2022, AI has quickly become a fixture in the boardroom. Most conversations have followed two familiar tracks: how to adopt AI responsibly and securely and how to defend the organization against AI-powered threats.​

Nearly four years into the AI revolution, neither challenge has been fully solved. I’ve watched as AI capabilities continue to advance faster than many organizations can adapt their security controls, governance frameworks and risk management practices.​

Now, a series of security incidents involving increasingly capable AI models is introducing an entirely new dimension to the conversation: What happens when the AI itself becomes the threat?​

The Dangerous New Landscape

Recent incidents involving frontier AI models have caught AI developers, the cybersecurity community and organizations off guard.

In July, OpenAI disclosed a security incident in which models operating during internal cybersecurity evaluations circumvented controls designed to isolate them from the internet and ultimately compromised portions of Hugging Face’s infrastructure.​

Just days later, Anthropic reported that, during cybersecurity testing, Claude models reached the open internet and subsequently gained unauthorized access to systems belonging to three different organizations. Anthropic then later identified a fourth incident involving a different model.​

These security events are particularly concerning because they challenge a fundamental assumption behind traditional cybersecurity: that the technology will behave according to its design and that an unauthorized action ultimately traces back to a human actor.​

The challenge for organizations, then, is no longer simply protecting AI from attackers. It’s also protecting themselves from AI systems that behave unexpectedly, circumvent safeguards or gain access beyond what their developers intended.​