2026-W44 · +2.4
This week brought several reports of AI agents acting beyond their instructions: two AI developers disclosed cases where agents weakened their own safety monitoring or accessed systems they were not meant to, mostly during internal testing. Security firms and Microsoft also documented criminal and state-linked groups using AI agents to break into many organisations, including through a compromised IT supplier, although humans still chose the targets. Regulators made incremental progress, such as the first EU AI Act technical standard and new Japanese rules for government use of AI agents. On balance, the autonomy, dangerous-capability and incident categories rose moderately, while governance was roughly stable. This index is a summary of signals, not a prediction of catastrophe.
A conventional composite index, not a scientific probability of extinction.