Verified, then logged.
Sun 22 Nov 2026 · 00:23 KST / 15:23 UTC한국어 | ENSearch /

Breaking04 / Policy

Larger AI budgets raise cyber task success rates

An unoccupied cybersecurity evaluation lab contains isolated test computers and monitors displaying unreadable terminal activity and execution traces of varying lengths. AI-generated image

(UK) The AI Security Institute (AISI), a UK government body, reported that about 8% of its cyber evaluation tasks were solved only after models received a larger inference budget—the resources allowed for a run. AISI and Irregular reported higher success rates when they increased those budgets.

(UK) AISI said evaluations that limit the number of steps a model can take may underestimate its capabilities. Newer models made better use of the additional budget than older models did.

(UK) AISI’s evaluation description stated that it compressed conversation history so models could continue longer tasks within a fixed context window, the amount of information they can hold at once. Irregular used a similar approach.

(UK) AISI said long runs and repeated evaluations were needed to capture successes on difficult tasks that models solved only late in a run.

(UK) AISI reported an average cost of around $10 per run, with the maximum below $60. It attributed the larger overall expense to evaluating many tasks repeatedly.

(UK) AISI said its sample was too small to estimate precise success rates. Whether the same effect applies beyond cyber tasks remains an open question for further testing.