Game mashups are among the harder things to ask a generative system for. An image only has to be plausible; a game has to be playable — rules that stay consistent, a loop that rewards the player, and a difficulty curve that does not collapse after five minutes. Melty generates playable mashups from prompts, and heise's hands-on test came away genuinely fascinated by the results (the full piece sits behind heise+).
That creative verdict is the less surprising half. What the test flags is security, and the word used is questionable rather than broken — the honest register for a product whose trust model is not yet public knowledge.
The security question is structural, not incidental. Tools that generate games or apps typically do three risky things at once: they execute code produced by a model, they host creations made by other users, and they accept unstructured natural-language input. Each is a known attack surface; together they interact. A shared gallery becomes a delivery channel for prompt injection, generated code runs with whatever privileges the platform granted it, and the boundary between content and executable blurs exactly where users are least equipped to see it.
There is a specific failure mode worth naming: a creative tool that executes what it generates inherits every flaw of the code it generates. The model does not need to be malicious; it only needs to be plausible. Plausible code with an unsanitised input path is what most real vulnerabilities look like.
This is now the recurring shape of consumer AI products. The demo is strong because demos reward novelty; the trust model is an afterthought because trust does not demo. AI app builders hit the same wall earlier this year, and the resolution was the same each time: sandbox execution, separate rendering from running, treat user-generated content as hostile by default, and say plainly what the platform can and cannot guarantee.
For anyone trying Melty, the practical rules are boring and effective: do not paste secrets or personal data into prompts, treat shared creations as untrusted content, and do not open a generated game in a browser session that is logged into anything you care about. For the category, the test is whether the security note arrives before the incident report.
