AI cyber task performance scope estimated to double every 4.7 months

The UK AI Security Institute reported that the length of cybersecurity tasks AI models can complete with 80% success has doubled roughly every 4.7 months since late 2024, based on its own test suite covering reverse engineering and web vulnerability exploitation. The measure reflects the human-expert time needed to complete a task, not the AI's actual runtime, and the institute cautioned that the finding does not mean the number or success rate of real-world cyberattacks is rising at the same pace. The estimate matters because it offers a benchmark for how quickly frontier models are absorbing complex security work, with implications for both defensive tooling and attacker capability. What to watch next is whether the roughly 4.7-month doubling trend holds across future AISI evaluations and whether it translates into measurable gains in live security operations.