agent security

An OWASP incubator project that protects AI agent memory from prompt injection, secret leakage, and tampering. It includes a Python library, policy-based controls, benchmarks, and integrations for agent frameworks like LangChain and AutoGen.

Bastiontrace is a Python library for analyzing AI agent tool-call traces to locate prompt injections, identify where they landed, and map the resulting blast radius.

Open-source authorization engine for AI agents that adds confidence-based gating, human review, policy-as-code, and audit logging. The repo shows quickstart code, local demo steps, SDK installs, and self-hosting options.

A hands-on kit for authorized red teaming of locally run LLMs, with jailbreak techniques, prompt-injection scenarios, streaming attack visualization, and support for OpenAI-compatible local model servers.

Open-source runtime guardrails for AI agents that intercept unsafe tool calls before they execute, aiming to block prompt-injection-driven exfiltration and other policy violations.

Open-source security research tool for auditing Python AI agents and codebases with AST analysis and Docker-based runtime verification.

Open-source input scanner for AI agents that strips hidden instructions and scans text, files, images, PDFs, QR codes, audio, and video before they reach an agent.

Local-first approval and receipt logging for AI agent sessions, with tamper-evident records and signed approvals for consequential actions.

A fail-closed execution firewall for AI agents that quarantines MCP tools, proxies OpenAI-compatible requests, and emits signed receipts for offline verification.

Open-source tooling for running REA, Ghidra, and JADX inside restricted Docker containers with an agent workflow for safer static analysis of untrusted files.

A local-first Rust runtime and IDE for running and governing AI agents with human approval, capability-based permissions, and audit logging.

Open-source agent-first display server for safely running human and AI-driven GUI sessions with per-app isolation and capability-scoped authorization.

An ICML 2026 research project page arguing that prompt injection comes from how LLMs misread roles, with an extended writeup, examples, and links to the paper, code, arXiv, and BibTeX.

VAIBot is an AI governance product focused on agent security. This page explains its prompt-injection threat model and positions the product as a circuit breaker that gates tool calls and other egress actions before they execute.

A research writeup from AgentSploit analyzing 1,384 tools across 28 MCP servers, highlighting how tool descriptions, source tools, and sink tools create an attack surface for agents.

Oconee Runtime is an enterprise AI governance product focused on policy enforcement for browser AI and coding agents. The page explains how it evaluates proposed agent actions against policy and supports allow, warn, and block outcomes.

A Reco security research article showing an AI-powered agent that maps Salesforce Experience Cloud sites, probes exposed objects and Apex methods, and attempts autonomous exploitation to find data exposure.