AI guardrails & agent security
3,625 agents · most starred first
AI guardrails & agent security, page 37 of 37
Surface for Semantic Kernel: the link-safety layer between an AI agent and the open web. One tool, surface_preflight, that tells your agent whether a link is worth fetching.
Surface for smolagents: the link-safety layer between an AI agent and the open web. One tool, surface_preflight, that tells your agent whether a link is worth fetching.
Surface for Haystack: the link-safety layer between an AI agent and the open web. One tool, surface_preflight, that tells your agent whether a link is worth fetching.
Tamper-evident audit trails for AI agent tool calls: SHA-256 hash chains and Merkle trees. Python port of the npm package @vluthra/agent-ledger, with identical hashes and roots across languages.
Compliance rail for AI agents - on-chain audit trails, AgentID soulbound identity verification, and EU AI Act Article 50 reports.
Open-source Apache-2.0 local runtime security for AI coding agents. Intercepts tool use (shell, secrets, MCP, installs) on your machine instead of trusting the model. Optional paid Guard Cloud.
A local-first trust, provenance, and review gate for AI agent tools, MCP servers, skills, and plugins.
Detect and rebut rationalization patterns in AI agent output
Vite plugin for LLM agent evasion: adversarial vocabulary, uglification, and comment injection.
Lightweight SDK for routing AI agent LLM traffic through a guardrails proxy
OKed for the OpenAI Agents SDK. Gates risky tool calls behind a human approval in Telegram or on the web dashboard.
Security for the Pi coding agent: zero-trust permission system, OS sandbox, tool guard, secret scanning and a Claude-style tool view. Pi permissions, Claude auto mode.
Pure, I/O-free kernel for agentic-pi: manifests, grants, policy PDP, audit hash chain + anchor, budget math, gate pipeline. Zero runtime dependencies.
Pi extension that turns a session into a governed agentic-org unit: default-deny, audited tool policy; bubblewrap-sandboxed bash; secret redaction; budget metering with escalation; org delegate/complete/escalate tools; manifest-listed project trust.
People, agents, and the access they share: delegation, invites, links, and an audit trail.
Signed agent mandates, constrained delegation and durable execution evidence.
Checks whether an agentic AI system's declared requirements are enforced in its code, with measured confidence and without any LLM.
Measure what an AI agent's sandbox lets it reach: network, secrets, host filesystem, privileges and processes, from inside the sandbox.
Fail-closed remote Curbi configured-rule checking for LangChain text runnables
Fail-closed remote Curbi configured-rule checking for completed LlamaIndex responses
LangChain tools and gates that check a SealKeeper SEAL before an agent calls an agent or an MCP server.
LangChain official tool integration for x402 CleanWeb Agent (87% token reduction, multimodal alt-text, anti-prompt injection, and autonomous micropayments)
Deprecated compatibility facade; use tba-provider-identity-jose instead.
AI agents send handup the real command, diff, file or email draft before they act. You approve, or deny with a note, from the terminal, desktop app or Android. Works with Claude Code, Codex, Cursor, omp and any MCP client. 14-day trial, then $30 once.
Agent Security Index is a free, independent directory of 102 AI agent security tools. It covers guardrails, red teaming, MCP security, observability and governance. Each listing has a plain summary, and the site sends a weekly briefing on agent security news.