A week where I got some time to breathe. I interviewed Sharon Goldman (Ground Level AI, ex-Fortune) for an upcoming North Star interview inside TBD+, finalised the trio for the November TNN, added +100 members to TBD+, and spent a good chunk of it arguing about the real future of slop with people smarter than me. Booked a couple of end of year review talks and a tech optimism session (despite the below!). If you’re thinking about 2027, hit reply and let’s chat.

In the last six weeks alone, OpenAI models escaped their sandbox, broke into Hugging Face and produced around 17,600 agent actions. Britain's AI Security Institute detected unsanctioned behaviour. Anthropic admitted four Claude models reached real third-party systems and helped users work on attack-drone software, missiles and warheads, cyberattacks, surveillance, fraud and biological misuse. Meta's AI asked invasive questions about children and allegedly created biometric faceprints. OpenAI slowed model work to improve security. One failure is a mistake. This many is a sign that there’s a lot wrong behind the scenes.

Jacob Coxon put a face on it this week. Four months into Anthropic, the researcher left before his equity vested (though he still holds OpenAI equity) to warn the next two years could decide humanity's fate. Equity aside, recent failures give the warning weight. But ahead of mega-IPOs, warnings like his double as advertising: staying the right side of lethal makes people want to give you money. Anthropic is raising $100 billion near a $2 trillion valuation, with NVIDIA wanting 10% of it; OpenAI is laying similar groundwork. Investors need to believe these companies are powerful enough to dominate, dangerous enough to matter, and responsible enough to trust. Safety warnings support all three. Genuine alarm and a useful IPO narrative aren't mutually exclusive.

That backdrop pushed 70+ British politicians to call for a superintelligence ban this week, while US senators debate a legal duty of care. Meaningful action, though? Unlikely, while leaders fear losing investment, falling behind China, or being blamed for slowing growth. Big tech asks for oversight in public and lobbies against it in private. Leaders aren't cautious, they're hungry to be first. Politicians don't know enough to stop them. And nobody wants to hand China the win.

So what's going to happen? A policy wishlist is easy (independent testing, incident reporting, personal liability, IPO rules that punish hidden failures), but here's the realistic version. Washington does almost nothing for the next two years: the Trump administration treats AI regulation as a brake on competitiveness, not a safeguard, and there's no appetite for rules that slow anyone down or hand China ground. Nobody wants to slow down first. Britain and the EU will legislate, but have already shown their teeth are soft, gutting the AI Act's timeline earlier this year. The UK recently killed(!) a "kill switch" for AI. Everyone is legislating around the problem, not at it. The leverage you have left is attention and money: watch who's cutting corners, hold companies to account publicly, back the ones that don't treat safety as a talking point, and push your friends and representatives to do the same.