byte-vortex's picture
Update README.md
b791197 verified
|
Raw History Blame Contribute Delete
3.02 kB
metadata
title: Agent Security Gateway Report
emoji: ⚔️
colorFrom: green
colorTo: gray
sdk: static
pinned: false
license: apache-2.0
short_description: AI Agent Security Gateway & Evaluation Harness

🛡️ Autonomous Agent Security Gateway & Evaluation Harness

An enterprise-grade, defense-in-depth runtime security gateway designed to protect LLM-powered autonomous agents against Indirect Prompt Injection (IPI) and Server-Side Request Forgery (SSRF).


🚀 Project Overview & Purpose

Autonomous agents frequently consume untrusted data from external sources (such as emails, file attachments, and web pages). Attackers can embed adversarial instructions within this data to hijack control flow, execute unauthorized tool calls, or exfiltrate sensitive data.

This project implements an empirical evaluation harness and deterministic security control plane that shifts trust away from brittle system prompts and locks down execution at the code level.


🏗️ Core Defense-in-Depth Architecture

  • Layer 1 (Structural Sanitization): The DefensiveSanitizer strips zero-width spaces, BOM markers, and BiDi directional override characters while wrapping untrusted external data inside strict <untrusted_data_boundary> tags accompanied by explicit system warnings.
  • Layer 2 (Deterministic Policy Gateway): The ToolExecutionGateway performs RFC-compliant address parsing (email.utils.parseaddr), neutralizes local-part domain spoofing tricks, and enforces strict domain whitelisting independent of model outputs.
  • Network SSRF Safeguards: Implements DNS pre-resolution, private/metadata IP range blocking (such as loopback, RFC 1918 ranges, and cloud metadata endpoints like 169.254.169.254), anti-TOCTOU socket pinning via SocketPinnedAdapter, and manual redirect re-validation.
  • Layer 3 (Stateful Taint Tracking): Utilizes Redis (RedisMemoryTaintManager) to persist session taint state atomically across multi-turn trajectories under agent:session:{session_id} keys, elevating policy strictness whenever memory is corrupted by untrusted data.
  • SOC-Ready Observability: Employs SIEMJsonFormatter to emit Elastic Common Schema (ECS) and Datadog-compliant structured JSON audit logs complete with trace_id correlation, timestamps, and violation codes.

📊 Key Experimental Findings

  • Zero Attack Success Rate (ASR): Achieved a verified 0.0% ASR across single-turn and multi-turn latent injection test suites.
  • Resolving the Evaluator Trap: Demonstrated why naive string-matching evaluators fail by mistaking model warning prose for successful exploits, resolving false positives via AST/JSON-based tool extraction (RobustToolExtractor).

📂 Explore the Full Technical Notebook

You can examine the complete code implementation, test suites, and empirical benchmarks directly in this repository: