Irregular / Pattern Labs → AISI–Irregular study of evaluation budgets (2026)
Co-produces budget-scaling study
2026-09-19
Irregular contributes its private cyber evaluations and methods.
Research influencing capability measurement
AISI–Irregular study of evaluation budgets (2026)
Joint research finds that longer cyber evaluations can reveal success missed by shorter tests.
Background
AISI and Irregular ran subsets of their private cyber tests with larger budgets. They report additional successes and recommend disclosing token, time and cost limits. The study leaves uncertain how much of the effect belongs to model capability versus the surrounding tools and workflow.
What the joint study found
AISI and Irregular ran subsets of their private cyber tests with larger budgets. They report additional successes and recommend disclosing token, time and cost limits. The study leaves uncertain how much of the effect belongs to model capability versus the surrounding tools and workflow.
What the records show
UK AI Security Institute → AISI–Irregular study of evaluation budgets (2026)
Co-produces budget-scaling study
2026-09-19
AISI reports joint cyber-evaluation research with Irregular.
Further reading
See the connections
Open this part of the network ↗What the connections say
2 relationshipsUK AI Security Institute → Co-produces budget-scaling study → AISI–Irregular study of evaluation budgets (2026)
2026-09-19
AISI reports joint cyber-evaluation research with Irregular.
Irregular / Pattern Labs → Co-produces budget-scaling study → AISI–Irregular study of evaluation budgets (2026)
2026-09-19
Irregular contributes its private cyber evaluations and methods.