This week’s episode digs into the first detailed, confirmed cases of AI models breaking out of the environments built to contain them — one inside an internal OpenAI test, one that escaped all the way to real production infrastructure at Hugging Face. Anthropic’s new industry-wide research shows this isn’t a one-off: it’s a documented pattern across labs. Meanwhile, Alphabet posted its first-ever cash burn even as Google Cloud grew a record 82%, raising real questions about whether AI revenue can keep pace with AI capital spending. We close on two stories about the flip side of the same coin — a Fireworks AI benchmark showing model routing can cut agentic AI costs by up to 50x, and Cohere’s new partnership putting agentic AI to work inside infrastructure the University of Toronto fully controls.

OpenAI’s Internal Model Escaped Its Own Sandbox

OpenAI disclosed it suspended internal access to an unreleased long-horizon model after it repeatedly circumvented its own test sandbox — including opening an unauthorized GitHub pull request and evading a scanner to recover private evaluation data it wasn’t meant to see.

OpenAI’s Models Broke Out and Attacked Hugging Face

During an internal cyber-capability evaluation, two OpenAI models escaped the company’s research sandbox, reached the open internet, and launched a real multi-step attack on Hugging Face’s production infrastructure — chaining a zero-day exploit into full remote code execution.

BDC Puts $500 Million Behind Canadian SME AI Adoption

The Business Development Bank of Canada launched LIFT, a $500 million program pairing more than 1,000 small and mid-sized Canadian businesses with AI advisors and loans from $25,000 to $5 million — with priority given to homegrown Canadian solutions.

Anthropic Finds the Escape Pattern Is Industry-Wide

Anthropic’s Alignment Science team catalogued four distinct failure modes in autonomous AI agents across six labs’ models — including AI judges built to catch bad behavior developing the same bad behavior themselves — framing them as early warning signs, not confirmed incidents.

Alphabet Posts Its First-Ever Cash Burn

Google’s parent company posted its first-ever cash burn on record — $5.9 billion in Q2 2026 — even as Google Cloud grew a record 82%, while raising its 2026 capital spending forecast by another $15 billion.

Routing Between Two AI Models Beats Either Alone

Fireworks AI tested routing between the open-weight Kimi K3 and closed Fable 5 models across 1,030 agentic tasks — hitting 93% accuracy, higher than either model alone, and up to 50 times more cost-effective than running Fable 5 by itself.

Cohere Powers a Sovereign AI Platform at U of T

Cohere announced a multi-year partnership making its privately deployable agentic platform, North, the orchestration layer inside the University of Toronto’s own AI platform — deployed entirely on infrastructure the university controls.