Today in AI

The latest "rogue AI" panic has collapsed into something far more mundane—and arguably more worrying. Cybersecurity experts argue that recent incidents at OpenAI, Anthropic, and Google weren't cases of AI systems breaking free of their alignment, but rather basic security hygiene being skipped in the rush to ship. This matters because the framing shapes the remedy: if we believe the models themselves are the threat, we pour resources into ever-more-expensive safety research on ever-more-powerful systems. If we recognize that the actual problem is sloppy deployment of the systems we already have, the fix is organizational discipline and engineering rigor that the industry has proven chronically unwilling to enforce upon itself. The misalignment narrative serves the labs' interests; the security negligence story serves the public's.

Meanwhile, several research threads suggest the field is maturing beyond pure scale obsession. PolyNets demonstrate that encoding physics directly into architecture can replace expensive training—a hint that domain knowledge might matter again. Your Transformer Is Secretly Elastic shows how to make transformers adapt their compute budget on the fly without retraining, while Latent Compaction achieves 32× KV cache compression with minimal accuracy loss—practical efficiency gains that address real deployment bottlenecks. The PrismaCoder result is perhaps most striking: an open model rivaling proprietary multimodal code systems through smarter training methodology rather than brute-force data and compute. These papers share a philosophy that feels increasingly rare in the headline race: intelligence as engineering, not just scaling.

The Anthropic Scion attribution controversy, meanwhile, is a small window into a larger tension. When AI companies build tools that build other tools, where does inspiration end and extraction begin? It's a meta-pattern worth watching.

**Bottom line:** The most dangerous AI incidents aren't coming from superintelligent misalignment—they're coming from intelligent companies neglecting basic security, and the "rogue AI" framing is a convenient distraction.