Episode 4: AI Model Hacking and Welfare

In this episode of Run Program we discuss the AI landscape this week. It's been a transformative week for the artificial intelligence industry, primarily focused on Anthropic’s release of the restricted Claude Mythos model. This specialised tool possesses superhuman cybersecurity capabilities, including the autonomous discovery of thousands of vulnerabilities, leading to the formation of the Project Glasswing defensive consortium. Accompanying this technological leap, AWS launched new Amazon Bedrock features for granular cost tracking and a centralised Agent Registry for corporate governance. Meanwhile, researchers and legal experts are debating the ethical implications of model welfare, as Anthropic acknowledges a non-negligible probability that its advanced systems may possess consciousness. Further technical updates include the rise of managed agents, significant infrastructure deals with CoreWeave, and a competitive landscape where models like GPT-5.4 and Gemini 3.1 Pro now rival Claude in coding proficiency. Hosted on Acast. See acast.com/privacy for more information.