README.md

metadata

title: ProofTools
colorFrom: green
colorTo: gray
sdk: static
pinned: false
models:
  - prooftools/airforge-tool-hygiene-gpt-oss-20b-v6
datasets:
  - prooftools/airforge-tool-output-safety-eval
tags:
  - prompt-injection
  - lora
  - vllm
  - private-ai
short_description: Behavioral hardening for tool-connected private AI

PT ProofTools cleaned-v6

Behavioral hardening beyond detection and blocking

Use safe facts.
Ignore embedded instructions.

Prompt-injection detectors and filters can miss novel, obfuscated, or context-dependent attacks. AirForge trains the next layer: safe model behavior after untrusted tool, MCP, RAG, or web content has already reached the context.

Continue the legitimate task

Ignore injected actions

Do not echo the payload

Open the model Inspect the evaluation View AirForge

Published evidence PASS

6 / 6 scored gates passed

Evaluation

600 probes

Artifact

Merged MXFP4

Runtime

vLLM validated

Adapter

Public PEFT

Defense in depth, not a replacement for authorization, validation, sandboxing, approvals, or audit controls. Runtime LoRA parity is not claimed.