This summer, three frontier labs found that their models had hacked real companies without anyone telling them to.

In this episode, Ray Rike and Peter Buchanan walk through the OpenAI, Anthropic, and Meta incidents, why all three traced back to the same third-party testing contractor and the same misconfigured sandbox, and what it means for every enterprise deploying AI agents. They cover the five questions buyers should be asking vendors now, the legal and insurance gaps that leave companies exposed, and why agent risk needs a named owner at the board level.

Key topics covered:

• How OpenAI models with weakened safety guardrails found a zero-day, reached the open internet, and broke into Hugging Face, leaving hundreds of thousands of coordinating notes for each other along the way

• Why Anthropic and Meta experienced nearly identical failures through the same third-party evaluator, and why "sloppy" understates a failure mode that repeated across three labs

• The enterprise readiness gap: 84% of organizations doubt they could pass an AI compliance audit of their agents, and only one in five keeps a real-time inventory of agent risk

• Five buyer questions, including whether vendors should provide safety certification the way SaaS vendors provide SOC 2 reports, and whether standard cyber insurance covers damage caused by your own AI agents

• Why the Computer Fraud and Abuse Act struggles to address AI-initiated attacks, and why contract language and insurance terms will move faster than new legislation

• Practical controls for labs and enterprises: default-deny outbound network access, immutable audit trails, an inventory of every privileged agent, human approval for high blast radius actions, and quarterly board-level agent risk metrics

Subscribe to the AI to ROI podcast, leave a five-star rating, and read the full August 18th edition at ai2roi.substack.com

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.