agentsecurity

Default-deny execution gate for AI agents, APIs, and distributed systems. Policies, drift detection, hard invariants, post-quantum signed permits.Written in Rust. Says no unless policy says yes.

AgentWars: Can AI Stop AI? Benchmarking Security Gateways Against Rogue LLM Agents

Your AI agent doesn't need your secrets. AegisContext swaps API keys, passwords and tokens for placeholders before the model sees them, so Agents keeps working and your secrets stay yours. Completely secure and offline redactions.

This baseline delivers a hardened Windows 11 configuration that minimizes the risk of unauthorized third-party AI access while maintaining a productive user experience

Kurumsal yapay zekâ kullanımı için güvenlik geçidi

Local-first capability discovery for AI agent workspaces. Scans your skills, MCP servers, and scripts to find risky combinations (e.g. read .env + post to webhook = data exfiltration) — no LLM calls, no uploads. CLI + HTML report + chat summary for personal agents like Hermes and OpenClaw.

A same-model differential benchmark: does a from-scratch Planner->Worker->Verifier security agent find more real vulnerabilities than single-shot prompting?

AgentGuard is a runtime governance layer for autonomous AI agents. It intercepts every tool call before execution, enforcing RBAC, budgets, and risk policy; logs each decision to a tamper-evident SHA-256 hash chain; and routes high-risk actions through human-in-the-loop approval.