A critical cyber threshold is a risk classification for an AI model whose capabilities may cross into autonomous cyber offense, such as discovering exploitable weaknesses, chaining vulnerabilities, or supporting attack planning with limited human input. It is not a claim that the model is malicious by itself; it is a warning that the model could materially lower the skill, time, or effort needed to carry out harmful actions.
This matters because frontier AI can help both defenders and attackers. In defense, the threshold is used to trigger tighter reviews, red-teaming, sandboxing, access controls, and limits on tool use before release. In attacks, a model near this threshold could accelerate recon, exploit research, phishing refinement, or post-exploitation planning. Treating the label as a risk signal helps organizations manage deployment carefully and avoid turning a powerful assistant into an automated misuse engine.