OpenAI tightens controls on its next model after warning it could reach “Critical” cybersecurity capability

OpenAI said it has tightened controls around a new model amid concerns about cybersecurity risks, warning that it could not rule out the system reaching a “Critical” capability level. In a statement covered by CNBC on August 10, the company described the potential threat as the ability to carry out cyberattacks targeting sophisticated cyber defenses, a classification that heightens the scrutiny applied to deployment decisions.

The company’s comments came as policymakers and security researchers increasingly focus on how advanced AI systems could be misused for hacking, vulnerability discovery, and automated exploitation. The debate has expanded beyond laboratory safety discussions to questions about accountability, guardrails, and what level of capability should trigger limits on release or use.

OpenAI did not frame the issue as a theoretical concern only about whether the model could be used safely, but rather as a gating question tied to the system’s demonstrated or potential capability. The “Critical” threshold referenced by CNBC is central to that approach, because it implies the model could meaningfully affect real-world defenses rather than merely produce low-impact or easily detectable outputs.

As part of the updated posture, OpenAI said it could not rule out that the model had reached “Critical” capability, indicating that the company’s risk assessment has moved toward more restrictive measures. The practical effect, according to the report, is that OpenAI is applying additional controls while it evaluates the possibility that the model could be used to target advanced protections.

The warning also underscores a broader pattern in AI oversight, where companies are increasingly required to document safety limitations and explain how they respond when systems approach or cross high-risk thresholds. For industry and public-sector buyers alike, questions about cybersecurity risk are directly tied to whether deployments should be delayed, limited, or subjected to additional verification steps.

OpenAI’s disclosure arrives during a period of heightened attention to AI security, including calls for clearer standards on how advanced capabilities should be assessed and what safeguards should be required before release. In that context, the company’s decision to emphasize uncertainty about capability levels suggests it is treating the cybersecurity risk as something that cannot be fully eliminated through assumptions alone.

What happens next depends on whether OpenAI’s internal assessments continue to indicate that the model remains below high-risk thresholds, and whether further mitigation measures can reduce the likelihood that the system could be used for high-impact cyberattacks. If the company continues to view the model as potentially at or near “Critical” capability, restrictions on availability and use are likely to remain a central component of its response.