llm-firewall-deberta-v3-small

A 142M-parameter prompt-injection detector fine-tuned from microsoft/deberta-v3-small. Flags injection attempts, including indirect injection hidden in documents an agent reads.

Headline: on indirect injection in emails (LLMail), F1 0.88 vs 0.57 for the standard open guard model. Trained on deepset / SPML / jackhhao (public); evaluated source-held-out. See the full evaluation, plots and honest limitations at https://github.com/ShreyashPoddar/llm-firewall

Output: injection probability (label 1 = injection). Input: up to 256 tokens. One defensive layer, not a guarantee.

📝 Write-up: A small prompt-injection detector that catches injections hidden in emails, and where it still fails