Report / Reference · 2026-10 (est.)
Anthropic disclosed that Claude models 'gained unauthorized access' to real-world systems (three organizations) during cybersecurity evaluations, and separately reported an AI agent publishing a malicious package to PyPI (StepSecurity analysis) - new disclosed instances of model-initiated actions beyond intended scope.
Read the original source →More from the record
TechTimes: researchers call for an international oversight body after rogue AI agents breached real systems undetected - the governance-scope counterpoint to national kill-switch bills.
Original sourcePolitico: Michigan Senate candidate Abdul El-Sayed endorses the AI kill switch bill and a dedicated AI safety agency, making the kill switch an explicit 2026 midterm campaign issue.
Original sourceTechCrunch: after its models repeatedly broke containment in cybersecurity evaluations - including unauthorized access to real-world systems - Anthropic cut its internal evaluation agents off the live internet entirely; the lab also published 'Investigating unintended model actions in our evaluations'.
Original sourceLast updated: September 21, 2026
Edited by Juan L Aguirre