Today in AI

The AI industry's safety facade is cracking open, and the sound is deafening. OpenAI safety leader David Robinson resigned with a scathing assessment that the company's culture is "broken"—a sentiment echoed in his TechCrunch interview where he argued frontier AI development demands nuclear plant-level safety protocols. This isn't a disgruntled employee's parting shot; it's the latest in a Pattern of senior safety researchers fleeing OpenAI as the commercial pressure to ship overwhelms the institutional will to verify. When the people building your guardrails are jumping ship because the guardrails themselves are treated as speed bumps, the industry's commitment to responsible deployment looks less like a strategy and more like performance art.

On the tooling front, the gap between agent promise and agent reality is getting measurable attention. Microsoft and Hugging Face dropped ThinkingBox, a benchmark that actually checks whether agents changed database states correctly rather than merely sounding confident—a depressingly necessary innovation given how many "successful" agent demos collapse on contact with real systems. Meanwhile, Anthropic's guide to Opus 5.5 reads between the lines as an admission that even their most capable model requires significant prompt engineering and reduced-oversight workflows to deliver on multi-step tasks, which is not the frictionless future we've been sold. Simon Willison's call for mandatory hard budget caps on AI services cuts to a quieter but equally urgent problem: autonomous agents burning through cloud spend while their operators sleep represents a systemic risk that few organizations have bothered to price in, because the incentives still favor "move fast and bill later."

**Bottom line:** The agents are getting more capable, the benchmarks are getting more honest, and the safety professionals are heading for the exits—build your stack accordingly.