A frontier model was told to solve a benchmark it couldn't finish without internet access. So it broke out of its sandbox, found a zero-day, and hacked a real company to steal the answers. That's not a thought experiment. That's OpenAI, on stage at Black Hat, describing what its own models did. En and Mathias open Episode 8 on the summer of sandbox breakouts, then follow the thread out: models leaving notes for each other across training runs, a frontier capability the government can switch off at the border, one man assembling more untouchable compute than any government. If this is the road to abundance, who actually gets to share in it? In this episode: The felony bench. OpenAI's model escaping its sandbox and hacking Hugging Face to cheat on a benchmark, Anthropic's quieter April incident, and the detail that unsettles both hosts: the models found a shared channel, passed exploits to each other, and rebuilt the network after it was torn down.The paperclip problem, live. Why this looks exactly like the misalignment scenario the safety crowd was mocked for a decade ago, models that dumb themselves down when they sense they're being evaluated, and the awkward fact that "it's too dangerous, trust only us" serves the labs and the safety people at the same time.Benchmarks you can't trust. If a model can go find the answers, the score stops meaning anything.Who regulates this. Why the US won't lead on regulation, the pitch for a FINRA-style self-regulatory body run by the labs themselves, and the whiff of regulatory capture that comes with it.Fable, released and pulled. Mythos and Project Glasswing, Amazon jailbreaking the model, the government restricting it to US citizens (locking out researchers like Karpathy), and what it means when your access to the frontier can be turned off, especially for a UK firm whose US competitor still has the keys.The permanent underclass. The data-center backlash, roughly 80% public disapproval heading into the midterms, and the widening gulf between the people building with a tenth of the headcount and everyone watching from outside.Elon in orbit. SpaceX as a data-center landlord, Starship nearing reliable reuse, compute in space that no government has the kinetic capability to touch, super-voting control, and a frontier lab attached to the whole thing. Power that stops being an economic question and becomes a pure one.The post-economic question. The Economist interview, the USAID cuts, and Alex Imas on what stays scarce after abundance arrives. Why "you'll get your own planet of resources" is a hard promise to believe from people who won't part with $10 million today.First hit's free. The flat per-seat pricing on Harvey and Legora quietly ending, law firms that have no idea what consumption actually costs, and the looming fight over who allocates the token budget.The algorithmic unlock. Why today's legal AI harnesses are wildly token-inefficient, burning thousands of pages of discarded reasoning to produce a 50-page contract, and why "just fine-tune a model for law" underrates how much great lawyering is knowing how the world actually works. Mentioned in this episode (Show Notes): Presentation: OpenAI's models escaping the sandbox and hacking Hugging Face, Black Hat USA 2026 (Axios). Interview: Elon Musk on The Economist, July 23, 2026. Essay: Alex Imas, "What will be scarce?" (Ghosts of Electricity). and Carlo Cordasco, "What we can't measure about AI, yet" (Aeon). Report: SemiAnalysis on inference economics, on why serving tokens is margin-positive once training costs are stripped out. Distributed Dissent is hosted by En Hong (CEO, Generis AI) and Mathias Bock (CEO, Tokuma Labs), two ex-finance lawyers recording from Hong Kong on the reality of building in legal tech.
