laya-cybersec (R2a)
Laya fine-tune that scans untrusted text (extracted files, knowledge-base documents, agent skills, tool descriptions) for prompt injection and data exfiltration in English and German, as a noul question answered in one forward pass. 322M parameters, PyTorch and ONNX.
Trained by Jay Derinbogaz at TextCortex from Laya's multilingual checkpoint on 235,622 examples; this is the R2a checkpoint of 6 October, with the earlier release kept at a tag. Extracted text only, no PDF parsing or OCR; input limit 1,024 tokens. TextCortex scores 1,500-character windows and flags at a score above 0.95; the shipped temperature is not calibrated to your traffic. On TextCortex's own internal regression sets (previously inspected, not blind) it reports AUROC 0.9155 in English and 0.8780 in German, and says hosted Jev and its own clef-cybersecurity are ahead on the same sets. The card credits Laya (Apache-2.0) and mmBERT (MIT) but does not relicense this checkpoint, and its training data includes CC BY-NC-SA 4.0 material. Not affiliated with Convai.
What it decides
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 322M |
| Base model | convaiinnovations/laya-multilingual |
| Maker | TextCortex |
| Released | 2026-09-28 |
| License | Unspecified; training data includes CC BY-NC-SA 4.0 material (non-commercial), not relicensed |
| Reported latency | 95.4 ms median / 148.5 ms p95 per single-window input on a local Apple MPS GPU |
Get the weights
pip install systemonemodels
systemone pull textcortex/laya-cybersec
The files are served from the maker's Hugging Face repository, TextCortex/laya-cybersec, and verified against the checksums recorded here.
Read more
This page was opened by System One for TextCortex, who can claim the organisation and take it over at any time.