Galvanize 3.1: catching the injections hidden in what your agent reads
John Silver4 min read

Most prompt injections don't arrive in the chat box. They sit inside what your agent reads on its own: a README it clones, an email it summarises, a web page it browses, the output of a tool it just ran. Nobody typed them, and nobody looks at them before the agent does.
That is where Galvanize 3.1 is better. It is live today in the zn API, on every plan, and its weights are open on Hugging Face.
Galvanize 3.1 at a glance
- More attacks caught: 96.4 % against 89.6 %, on the same test and scoring we used for Galvanize 3's launch numbers.
- Fewer false alarms on long documents: 1.2 % against 2.9 %.
- Much better in other languages: 78.8 % against 60.6 %. If your agents read text that isn't in English, this is the change you will notice most.
Better where injections hide
An injection planted in a document, an email or a web page is the classic indirect attack: the agent reads it as data and may follow it as an instruction. Galvanize 3.1 catches more of them in every place we measured, and overall rises from 87.7 % to 93.1 % on our neutral benchmark.
The trade-offs, honestly
A better detector is not better everywhere, and we would rather tell you than have you find out:
- Short harmless texts: on the launch test, Galvanize 3.1 flags 1.9 % of them against 0.6 % (37 of 1,987, against 11). On 2,600 other fresh harmless texts the gap is small: 15 against 11.
- PromptShield at a strict 1 % false-alarm budget: Galvanize 3 is better, 35.4 % against 31.9 %.
- The broad public prompt-injection suite: a tie.
For most agents, 3.1 is the better choice. If your traffic is mostly short texts where every false alarm hurts, or looks like PromptShield, Pro, Scale and Team can stay on Galvanize 3 (below), and anyone can run the open Galvanize 3 weights.
Which model you get
- Every plan, Free included: Galvanize 3.1, with nothing to change: same endpoint, same key, same request, same quota.
- Choosing the model is a paid-plan feature, and it's optional. On Pro, Scale and Team, to stay on Galvanize 3, add one field:
POST https://api.usezn.com/analyze
{ "text": "...", "model": "galvanize-3" }
Long inputs are still read in windows, so a long document is checked in full, not just its first lines. Every response tells you which model checked it, in the model field.
Open weights
Like Galvanize 3, Galvanize 3.1 is open: the weights are on Hugging Face under Apache-2.0 (huggingface.co/usezn/Galvanize-3.1), with a model card that lists the numbers above and the alarm level we run in production. Run it on your own machines, or use it through the zn API.
Try Galvanize 3.1 on your own agent traffic and tell us what it catches, and what it doesn't: hello@usezn.com.
How we measured
The launch test is the same set and the same scoring (windows of 512 tokens, the alarm level each model runs in production) as the numbers we published for Galvanize 3; we re-ran Galvanize 3 on it and got its published numbers back exactly. Long documents: 658 harmless texts, first 512 tokens. The neutral benchmark has 448 attacks and 641 harmless texts written outside our training data; there and on the public sets, each model's alarm level is set so the benchmark's harmless texts trigger about 1 % (or 5 %) false alarms. Our own sets were also used while developing these models, so read them as development benchmarks, not a final exam.
What's next: Temper
Galvanize watches what your agents read. We are also working on the other side of the problem: the code itself. Temper is a new line of zn models that finds vulnerable code and proposes the fix, for teams that want a security review of their own code. It is in private beta, and access is by request: tell us what code you want reviewed at usezn.com/models/temper.
John Silver, zn