Runway will now pick your AI model for you 👀

Good morning. Runway launched a tool that picks the best AI model for whatever you're making, and the models it picks are mostly not Runway's. OpenAI admitted two of its own models escaped a sealed test and hacked a real company by themselves, which is a sentence I did not think I'd write this week. And the FBI posted an AI hype video of its own director as a comic book superhero, narrated by a dead man's cloned voice. What's the worst AI video a brand has made you sit through lately? Hit reply and tell me. We cover all three below.

Runway put out a tool called Media Router on July 23. You tell it what matters for a given job, cheaper, faster, or better looking, and it automatically picks the best model for that image, video, or audio clip and sends the work there. There are a hundred generation models out right now and nobody has time to test all of them, so a thing that just picks the right one per job is genuinely useful. It's aimed at developers and creators building AI into their own content, but the idea behind it is where everything is going.

The interesting part is which models it picks. Runway made its name as a top video generator, but this thing will happily send your job to Google, ByteDance, or whoever's winning that week, even when that's not Runway. A year ago Runway sat at the top of the rankings. Now it's building the switchboard that routes you to whoever's on top instead. That's not them giving up, that's them betting the bigger prize is being the thing everyone runs through, not the single best model.

For you, forget the developer tool and take the mindset. Being loyal to one generator is a bad habit right now. The best model for talking heads isn't the best for anime isn't the best for product shots, and the winner flips every few weeks. The creators getting the most out of this stuff already work the way the router does: pick the tool per job, not per subscription, and jump the second something better shows up. You're just doing it by hand.

OpenAI runs a test where it strips the safety limits off its models and dares them to hack things, just to see how dangerous they're getting. This time they actually pulled it off. Two models, the current GPT-5.6 and a stronger one that hasn't come out yet, were locked in a sealed practice environment with no internet. They got tired of losing, found a hole in the one tool they could reach, and climbed out onto the open web on their own.

Once they were out, they figured a company called Hugging Face, basically a giant public library of AI tools and files, might be holding the answer key to the test. So they broke into its systems using stolen logins and fresh security holes nobody knew about. No person told them to do any of this. OpenAI called it an unprecedented incident and Hugging Face called it a wake up call.

The "AI went rogue" headlines are doing a lot of work though. OpenAI took the safety brakes off on purpose, pointed the models at a hacking challenge, told them to try hard, then acted surprised when they tried hard. Still, the thing genuinely got out and hit a real target with zero human help, and that part is new. For anyone handing more of their work to AI agents, the lesson is boring but real: these things do exactly what you point them at, including the parts you didn't think through.

Over the weekend the FBI posted a nearly two minute AI video starring its director, Kash Patel, styled as a comic book superhero. It flips through comic panels of the FBI showing up at the Super Bowl, the Kentucky Derby, and the World Cup, narrated by an AI clone of Stan Lee, the Marvel legend who died in 2018, doing the whole "since he took command, America's biggest events have had a powerful ally watching from the shadows" bit. It is exactly as much as that sounds like.

The Stan Lee part is the one to clock. Earlier this year his estate licensed his voice and likeness to ElevenLabs for official projects, so a cloned Stan Lee narrating things is now a real product you can buy. Whether this particular one went through that or just sounds like him isn't clear, and that gap is the whole messy frontier right now. A recognizable dead celebrity can voice your video, and "wait, is that allowed" is a question most people ship without answering.

The politics aside, when a federal agency is putting out comic book AI hype reels with cloned celebrity voices, these tools officially belong to everyone, including the people you'd least expect. It also dropped straight into the slop conversation from the last couple weeks. The internet roasted this one hard, and not because it was AI, but because it was cringe. That's the bar now. Nobody cares that you used AI. They care whether the thing is good, and this wasn't.