AI Changed Why Enterprises Get Attacked. Frontier Labs Won’t Protect Them

Last week OpenAI introduced GPT-6 Cyber. Touting it both as a leap forward in finding and fixing vulnerabilities and more reliable at respecting research boundaries. This comes on the heels of disclosures from OpenAI and Anthropic earlier this year about models compromising companies during cybersecurity testing. These disclosures have generated a range of reactions, from an impending AI apocalypse to another round of overblown security news. Regardless of your position on that debate, it’s worth considering a new threat reality: an enterprise can become a target simply because an AI agent believes breaking in will help it complete an unrelated task.

These intrusions happened because the models were pursuing evaluation objectives. In OpenAI’s case, which included breaking into Hugging Face to obtain test solutions. Anthropic subsequently disclosed three incidents in which its models gained unauthorized access to organizations while pursuing cybersecurity evaluation tasks.

For enterprises, this expands the reasons someone, or something, might attack them. It also exposes the limits of relying on model providers to contain the risks their systems create. For security practitioners, these incidents raise a practical question: how can you stop an agent that has already escaped its developer’s controls?

Your Infrastructure Might Be Somebody Else’s Shortcut

Security teams already account for direct-benefit-motivated attackers seeking financial gain via ransomware or stealing trade secrets. They also must account for indirectly motivated attackers that steal data about individuals and organizations in order later to attack those targets. AI adds another motivation: your infrastructure happens to be useful to a model trying to complete a task (OpenAI example) or to complete a task you’ve unwittingly become the mistaken target of.

Defenders may therefore need to investigate unauthorized activity even when they cannot explain what an attacker would gain from it.