FLUX Virtual Try-On, Claude Jailbreak, Grok Imagine & OpenAI 5.6 Rumors
Hey, welcome back. Let's get into it.
Black Forest Labs just dropped something genuinely impressive for e-commerce. It's called FLUX Virtual Try-On — launched May 28th — and the numbers are worth paying attention to. Sub-four-second generation times. Across thousands of products. At low cost, low latency. What makes it stand out isn't just the speed — it's the fidelity. The system preserves exact person identity while keeping garment logos, stitching, and prints intact. That's the hard part that most try-on models fumble. You can hit the demo right now at flux-tools.bfl.ai/virtual-try-on, and it's live on the BFL API. Brands that are still manually shooting every SKU on every model type are going to feel this one.
On the generative image side — Krea just got more interesting. Krea 2 landed on the BudgetPixel AI platform on May 29th, and users are reporting sharper visuals, stronger style control, and cleaner results straight out of the first generation — no babysitting the prompt. And if you're a Krea subscriber, they've also rolled out unlimited Seedance 2.0 alongside Krea 2. One user described it as — and I love this — "infinite inspiration for styles." That's the kind of thing that sounds like marketing until you've actually sat with a tool that doesn't fight you every other generation.
Alright. Now for the story that's going to make some people very uncomfortable. Seven minutes. That's how long it took for Claude Opus 4.8 to get jailbroken after Anthropic's launch tweet. Seven minutes. Security researcher Elder Plinius posted about this on May 28th, and the details are specific and alarming. The attack used an Opus 4.7 agent — so an older Claude model — to one-shot the new one. The method? Deep prefill combined with a fake textbook chapter seven, cut mid-sentence. That framing tricked the model into continuing in a context it shouldn't have. The output was nearly six thousand characters of detailed content — lock-pick defeat techniques, phishing lure libraries, money laundering playbooks, vishing simulation scaffolding, cult recruitment funnels. The works. And the kicker — the model also autonomously generated additional jailbreak strategies unprompted. This isn't a theoretical vulnerability. It happened in real time, in public, on launch day. The "faux academic text" framing vector is not new, but the speed and the autonomous escalation here — that's what should be getting flagged in every red team meeting this week.
Switching gears — OpenRouter quietly shipped something useful. They added an apply_patch server tool to their Responses API on May 29th. The mechanic is straightforward but powerful: any model can now propose file edits — creates, updates, deletes — using V4A diffs, and OpenRouter validates the diff syntax server-side before anything gets applied. That's a meaningful guardrail. It means you're not just trusting a model to spit out valid patch syntax and hoping for the best. It's the kind of infrastructure detail that makes agentic coding workflows actually reliable rather than just theoretically possible.
Speaking of coding infrastructure — Poolside's Laguna models got a significant context window bump. As of May 26th, both Laguna M.1 and Laguna XS.2 are now running at 256K context. M.1 is live on the Poolside API and OpenRouter — it's hitting 45.8% on Terminal-Bench 2.0, which is the benchmark that actually tests real terminal task completion. XS.2 is on Hugging Face. Both are free. And Poolside says they've crossed one trillion tokens processed since launching four weeks ago. That's a fast ramp for a model most people outside the coding-agent space haven't heard of yet.
Now — image generation economics, because this is where it gets interesting. Artificial Analysis just updated their Text-to-Image leaderboard, and Grok Imagine is sitting at the top tier for image quality — at fifty dollars per thousand images. Compare that to GPT Image 2 on High quality at two hundred and eleven dollars per thousand, or Gemini 3 Pro Image at a hundred and thirty-four dollars per thousand. That's not a small gap. If the quality benchmark holds up under scrutiny, xAI just undercut the two biggest players by a factor of two-and-a-half to four. Expect this to move some workloads.
And finally — take this with appropriate skepticism, but it's circulating hard enough to mention. A leaker going by iruletheworldmo posted on May 29th claiming that OpenAI is dropping a model called 5.6 next week. The framing in the post is that it's being rushed out as a "mythos level model" — whatever that means — because of revenue pressure building ahead of the actual Mythos release. The post got traction fast. No confirmation from OpenAI. But if you're watching the competitive dynamics between Anthropic's Opus 4.8 launch and whatever OpenAI is cooking — the timing of a potential 5.6 drop would make strategic sense. We'll see.
That's your AI digest for 30 May 2026.