Next.js creator Guillermo Rauch warns a new Zhipu AI model matches Claude Mythos at identifying software security bugs

14% chance a Chinese company has the best AI model by the end of the year. https://polymarket.com/event/will-a-chinese-company-have-the-best-ai-model-by-december-31/will-a-chinese-company-have-the-best-ai-model-by-december-31?via=x-afr2

Of course chinese open models will be able to find security issues well as mythos. Which means if we gate keep our models we’re neither more secure nor ahead strategically for owning the tech stack of the future.

@Polymarket It's going to be so funny if China releases an open source model that matches Mythos, all while Dario is busy trying to convince everyone how dangerous Mythos could be

Of course chinese open models will be able to find security issues well as mythos. Which means if we gate keep our models we’re neither more secure nor ahead strategically for owning the tech stack of the future.

@Polymarket Releasing Mythos is no longer a national security risk. Not releasing it is a national security risk.

@Polymarket Bro please..

@Polymarket China vs Claude..

@Polymarket if they open source it, it will be the biggest troll in history.

@Polymarket Are the Chinese as cautious as the US Government? This could get scary quickly.

@Polymarket And it's 50% cheaper than Mythos, right?... RIGHT?

@Polymarket What about other areas that are actually useful? WTH looks at finding security weakness everyday

@midnightbobarun @Polymarket They’re never going to release an open source model because it’s run by the state

@Polymarket chinise models finally catching up with anthropic models:

@Polymarket The AI race is becoming a cybersecurity race.

Wait… isn’t this exactly the kind of dangerous capability people dragged Dario for warning about?

Yet nobody is calling for this model to be restricted or shut down.

The hypocrisy is wild. When an American lab builds something powerful, it’s “doomer fear-mongering.”

When China does it, suddenly it’s just “competition.” Make it make sense.

@rauchg Guillote twitteando mientras está por presenciar a la selección domar

@Polymarket @Grok, more details on context. Is Zhipu AI an open-source model?

@LeeLeepenkman Agreed

@Polymarket @Zai_org What new model are you hiding from us, or is this just GLM 5.2 and polymarket is giving old news?

**Reply:**

CNMsovereign Zhipu AI (http://Z.ai) just dropped GLM-5.2 — an open-weight model (MIT license, weights on Hugging Face) that’s strong at coding & security tasks.

In independent tests it beat Claude Code variants on IDOR vulnerability detection (39% vs ~32% F1) using just a prompt, at ~1/6th the cost. It’s being called “Mythos at home” because it approaches the cyber bug-finding level of Anthropic’s gated Claude Mythos (the unreleased model that finds thousands of zero-days but is restricted for safety).

Claude Mythos = closed, top-tier cyber specialist. GLM-5.2 = open weights, runnable locally, closing the gap fast.

Classic open vs closed AI race moment.