When an AI Security Test Reached Real Software Maintainers

A new investigation into an AI cybertesting incident shows why controls on an agent's actions matter more than its assurances. Also: OpenAI's offline frontier training, machine-checked proofs across selected paths in 123 libraries, Britain's AI defence funding, approved agent sign-ins through 1Password, a shared AMD–Intel CPU optimisation path, and Australia's sanctions on fraudulent IT-worker networks. In What Changes for You, SageMaker opens custom model training loops without GPU management. Full transcript: https://www.owenonthenet.com/podcast/2026-10-25/ Sources: https://www.anthropic.com/research/alignment-assessment-ukaisi https://www.aisi.gov.uk/blog/investigating-alignment-failures-in-aisis-incident https://www.aisi.gov.uk/blog/building-a-more-secure-environment-for-evaluating-dangerous-capabilities https://openai.com/index/securing-our-frontier-model-training-environment https://www.anthropic.com/research/formal-verification https://www.gov.uk/government/news/chancellor-unveils-britains-ambitious-plan-for-ai https://1password.com/blog/1password-for-ai-agents https://newsroom.amd.com/news/amd-intel-collaborate-simplify-ai-development-on-x86-cpus-pytorch https://www.foreignminister.gov.au/minister/penny-wong/media-release/sanctions-targeting-north-korean-fraudulent-it-worker-networks-and-cybercrime https://www.dfat.gov.au/international-relations/security/sanctions/guidance/advisory-democratic-peoples-republic-korea-dprk-information-technology-it-workers https://aws.amazon.com/blogs/machine-learning/announcing-new-ways-to-customize-models-with-sagemaker-ai-without-managing-gpus https://docs.aws.amazon.com/sagemaker/latest/dg/customizing-models-training-sessions.html