Overview
Severity: MEDIUM | Affected: Aegis AI | Category: tool
AI security firm Aegis AI has launched GuardianML, an open-source security framework designed to act as a firewall for applications built on large language models. The tool is intended to be deployed as a middleware layer, providing real-time scanning of both user inputs and model-generated outputs. For inputs, GuardianML detects common attack patterns like prompt injection, jailbreak attempts, and the presence of personally identifiable information (PII). For outputs, it scans for data leakage, toxic content, and responses that violate predefined security policies. The framework leverages a combination of deterministic rules, heuristic analysis, and a lightweight, specialized classification model to ensure low-latency performance. The release aims to standardize security practices for LLM implementation, offering developers a ready-to-use solution to mitigate risks associated with direct model interaction and promote the adoption of LLM-SecOps principles.