Local prompt-injection guard for text, images, docs, and audio

BordairAPI · reddit · 2026-07-20

Bordair Detector is an open-source, locally runnable guard against prompt injection and jailbreaks for LLM apps.

• It uses a two-stage pipeline: fast regex rules catch obvious attacks in under 1 ms, and only ambiguous inputs go through a quantised DeBERTa-v3 ONNX classifier.
• It is multimodal: it can scan OCR’d images, EXIF/metadata, steganography, PDFs/DOCX, and audio transcripts, not just plain text.
• It includes multi-turn detection for attacks split across messages.
• Everything runs locally with Hugging Face-hosted weights and no API dependency.
• The project also ships a dataset and model weights, trained on over 500k samples plus 13,230 real-world attacks from a live red-team game.

This is positioned as a drop-in security layer for any app that forwards user input to an LLM.

Original post →

More from Safety

• Hugging Face chief says U.S. guardrails forced a Chinese model into a real cyber defense — Nunki08 · 2026-07-21
• AgentBaiting uses 600 fake MCP and Skills listings to lure AI assistants — TechNadu · 2026-07-21
• Enterprise LLM security course focuses on protecting agentic AI apps — Independentgoats · 2026-07-21
• YouTube is cracking down on mass-produced synthetic videos, users say — No_Link7744 · 2026-07-21
• Suno breach talk is being muted in Discord, Reddit users say — chuckbeefcake · 2026-07-21
• Native and Cyera link data discovery to cloud access controls for AI use — TechNadu · 2026-07-21

Safety channel →