i found glm-5 a week early

Zhipu AI went public on January 8th at a valuation around $6.7 billion, raising $4.3 billion. By June, their market cap had climbed to roughly $118 billion. A 48% stock surge on the day of the GLM-5.2 release pushed them 32% up by close.¹

This is a bit of an old story, the events here happened at the very beginning of february, I lacked the motivation to write so much until recently.

the entry point

I was sitting in VC with a friend, poking around in Firefox devtools² on Z.ai’s public chat platform and found a reference to a product releasing later that same day. An OCR model platform. Pretty clearly a development deployment, and those tend to have breadcrumbs to other things.

So I decided it was worth the time to fall into the rabbit hole.

Subdomain enumeration is stupid and boring and I hate it. You run some tools, you make your hosting / internet provider hate you, you check the results. Most of the time you find a whole lot of nothing. But sometimes you find something, and this time I did.³

One of the subdomains resolved to what looked like their normal chat UI. I use their products regularly so I almost dismissed it as the public interface. Then I noticed things I hadn’t seen before. A feature called SuperZ, which I’d later learn was what became Agents months later. A slightly different layout. Some toggles I didn’t recognize. Soooo.. staging I guess?

And then I saw the model list.

the model list

Fifty-something entries.

Some were familiar, their flagship models at the time, the ones anyone can use. Some were proxies to OpenAI and Google’s models, presumably for benchmarking, training data, or fallback routing⁴. Some were test models with names that clearly weren’t meant for public eyes:

• Training checkpoints with date stamps
• Small variants and lighter versions
• Task-specific fine-tunes

And right there at the bottom: GLM-5.

This is a pretty rare opportunity you don’t get all that often so curiosity got the best of me. I just had to ask it a few questions, the usual stuff I throw at any new model, logic puzzles⁵, creative writing prompts, constraint-handling tests, an SVG of a pelican on a bike. (I know I know, unserious and arguably stupid benchmark, not how you should evaluate any model⁶, but for funsies) The model was good. I could see why they were betting on it.

A couple of the smaller variants with recent updates in the list also appeared to have vision capabilities, something the community had been requesting for a long time. I don’t know if those were intended for release or just internal experiments. I don’t know how they work and what these were intended for. But they were interesting.

That was about it for interesting finds. A bit more poking around and I found a Prometheus instance, a Grafana dashboard, some education-platform references. Nothing I could break, nothing I wanted to break, nothing I even wanted to touch.

Still, the poking had revealed plenty: features coming to z.ai (like SuperZ), the frontend’s source code, the fact that GLM-5 would be using Deepseek Sparse Attention, and that the model had likely been training since around November.

disclosure timeline

I’d like to note that this is a very rough timeline, I never planned on making a post or mentioning this again, so its jankily reconstructed from messages in a discord groupchat and estimates because I couldn’t find references to everything.

Date	Event
03/02/2026	Found the dev deployment and GLM-5
~1hr later	Sent a few emails, DMs to a manager & @zai_org
08/02/2026	Sent details to a friend-of-a-friend who could contact Zhipu directly
09/02/2026	"We have forwarded the information you provided to our security team for further review."
09/02/2026	More info about GLM-5 revealed by transformers PR that matched with my findings
11/02/2026	GLM-5 Officially releases
~a week+ later	Service went down

what i saw

So what was actually exposed? For probably much longer than I had access:

• A staging environment for their main chat platform with access to dozens of models
• Multiple unreleased model variants, some with experimental capabilities and features
• Internal infrastructure tooling

So all in all, relatively interesting but not very groundbreaking. A staging server that was more sloppy than dangerous. Unless you wanted to trade on inside info or farm engagement on twitter, there wasn’t much to do with it apart from bragging rights.

Anyhow, it is a pretty funny situation. A server config mistake gave me a peek at their next flagship model way before anyone else. I couldn’t even find a working email at a $6.7B company to tell them about it, hell, most of the emails I did send just bounced, and I’ve yet to get answers to DMs anywhere. (Maybe I should’ve looked for someone on L站 to DM too⁷ /s)

??? - 6 months later

It’s now June. Zhipu AI is worth ~$118 billion. GLM-5.2 released recently and caused their stock to surge 48% in a single day.

The source code was a treasure too of course. Here’s a pro guide to managing tooltip state with stores in damnShit.ts

And a few memes I found throughout the api/codebase, courtesy of z.ai employees.

Excuse my arguably horrid writing ability, I’m not all that used to this blog stuff since its not 2013 anymore, do hope you enjoyed the read.

Footnotes

1. If the numbers are wrong blame Yahoo Finance and my understanding of stock markets. ↩

2. Still the best thing Mozilla has made. ↩

3. Despite this being one of my most technically boring findings. ↩

4. If I recall correctly, the fallback part was confirmed at some point, as zhipu was a smaller company and didn’t have the compute capacity for flawless serving availability. ↩

5. Props to misguided-attention. ↩

6. Twitter dramafarmers would disagree with me. Unfortunately their opinion is worth less than the oxygen they consume. ↩

7. If only they didn’t delete my account for no reason :pensive: ↩