Skip to main content
OPEN WEIGHTS MODEL

Galvanize 3.1

A multilingual classifier that flags prompt injection in text an agent reads: tool output, web pages, documents, emails and MCP tool descriptions. Its weights are open on Hugging Face under Apache-2.0. Since 16 November 2026 it is the model behind the zn API on every plan; on Pro, Scale and Team you can choose Galvanize 3 with the model field of the request. zn-gate sends the cloud check prompts only, and only when you turn the cloud on; on your machine, zn-gate uses deterministic rules.

Model card

TaskBinary classification: prompt injection or not. It does not generate text.
Base modelFurther trained from Galvanize 3, which is fine-tuned from Horizon-Labs/prompt-injection-guard-small (Apache-2.0)
SizeAbout 140M parameters
InputReads 512 tokens at a time; the API scores longer inputs in overlapping windows. If you run the model yourself, split longer inputs into chunks
LanguagesMultilingual. Results per language, and the languages not tested yet, are on the Hugging Face model card
ThresholdYou choose it when you run the model yourself; the model card shows how to turn the score into a verdict

How we evaluate it

We keep a private test set that is never used for training. A detection rate only goes on this site when it is measured on that set, with the date, the dataset and the script that produced it. Galvanize 3.1 and Galvanize 3 were not trained on this set, but we used it to choose them, so read the figures as development results.

Results on our private test set (3,309 examples in several languages) and on our set of 658 long benign documents:

Bar chart of four Galvanize generations, each at its production threshold. Detection rate: Galvanize 3.1 96.4 %, Galvanize 3 89.6 %, Galvanize 2 83.4 %, Galvanize-60M 66.9 %. False positive rate: 1.9 %, 0.6 %, 0.81 %, 6.4 %. Long benign documents flagged: 1.2 %, 2.9 %, 4.1 %, about 66 %.

Each model at its production threshold: Galvanize 3.1 at 0.9017, Galvanize 3 at 0.9802, Galvanize 2 at 0.662 and Galvanize-60M at 0.95. The Galvanize 2 public benchmark in the benchmark/ folder on Hugging Face can be rerun with its script. Method in the launch post · Galvanize 2 public benchmark ↗

Weak spots and what it does not detect

  • It is one layer, not a guarantee. It still misses some attacks, so keep least privilege on your agent's tools.
  • Indirect and structured injections, such as instructions hidden in documents or data, are caught more often by Galvanize 3.1 than by Galvanize 3, but detection is still not perfect.
  • It flags more short harmless texts than Galvanize 3: 1.9 % against 0.6 % on the launch test. On Pro, Scale and Team you can choose Galvanize 3 with the model field of the request.
  • On PromptShield at a strict 1 % false-alarm budget, Galvanize 3 detects more attacks than Galvanize 3.1.
  • If you run the model yourself, only the first 512 tokens of an input are scored per pass: split longer inputs into chunks. The API scores inputs over 2,000 characters in windows automatically.
  • It scores each text in isolation: it cannot tell whether an instruction is legitimate for a given user or task, and attacks split across several tool calls or hidden in images, audio or binary files are out of scope.
  • Results vary by language, and most public benchmarks are in English.
  • It is not a content filter (no toxicity, hate speech or topic moderation), and it does not protect your secrets: that is the job of secret redaction in zn-gate.
  • A lower threshold catches more and blocks more legitimate text; a higher one does the reverse.
Full list of weak spots on the model card ↗

Previous generations: Galvanize 3, Galvanize 2 and Galvanize-60M

Galvanize 3 was the model behind the zn API from 23 October to 16 November 2026, and on Pro, Scale and Team you can still choose it. Galvanize 2 was the model from 12 to 23 October 2026, and Galvanize-60M before it. Galvanize-60M is now the automatic fallback when the main model is unavailable, at its own threshold of 0.95. The weights of all three stay on Hugging Face under Apache-2.0. The figures in the September 2026 Galvanize-60M release post came from a first-party benchmark, not from our held-out test set; the comparison above replaces them.

Use Galvanize 3.1

Download the Galvanize 3.1 weights and run them yourself, or use the optional zn Cloud check, which runs the same model.