threatdeck
Live 13:15 UTC ·

Tool radar

Agentic AI tools that attack and defend, tracked from their repositories. Offensive tools are listed so defenders know what they face: each one carries defender notes, never attack instructions.

95 tools published: 46 reviewed by the editor, 49 defensive tools found automatically and not yet reviewed. Offensive and dual-use tools are published only after review.

Seen in attacks

Editor-verified reports of use in real intrusions
  • ARTEX Offensive agent · Offense · code withdrawn · 12 stories

    Reported in real intrusions against hundreds of organisations: review vendor reporting (CrowdStrike, Microsoft, Google GTIG, Wiz) for infrastructure and TTPs, hunt for bursts of automated recon and login probing against internet-facing apps, and patch exposed services quickly. Do not run untrusted forks.

Tool radar: 95 tools by family and adoption ring AdoptTrialAssessHold Offensive agents & red-teamingGuardrails & agent securityModel & code securityAI SOC & benchmarks Adrian · Agent & MCP security · Assess Adrian Adversarial Robustness Toolbox · LLM red-teaming · Trial Adversarial Robustness Toolbox agent-governance-toolkit · Agent & MCP security · Assess agent-governance-toolkit agent-safe-pipeline · Agent & MCP security · Assess agent-safe-pipeline agent-safehouse · Agent & MCP security · Assess agent-safehouse agent-threat-rules · Agent & MCP security · Assess agent-threat-rules AgentDojo · Benchmark · Trial AgentDojo Agentic Radar · Agent & MCP security · Hold Agentic Radar AgentZ · Agent & MCP security · Assess AgentZ ai-exploits · Model security · Hold ai-exploits AiSOC · AI SOC · Assess AiSOC akto · Agent & MCP security · Assess akto arcjet-js · Guardrail · Assess arcjet-js Argus-Agentic-SOC-Platform · AI SOC · Assess Argus-Agentic-SOC-Platform ARTEX · Offensive agent · Hold ARTEX Augustus · LLM red-teaming · Assess Augustus beelzebub · Benchmark · Assess beelzebub bernstein · Agent & MCP security · Assess bernstein bla · Agent & MCP security · Assess bla BoxPwnr · Benchmark · Assess BoxPwnr CAI (Cybersecurity AI) · Offensive agent · Hold CAI (Cybersecurity AI) Casco · Offensive agent · Assess Casco cc-safety-net · Agent & MCP security · Assess cc-safety-net Cisco MCP Scanner · Agent & MCP security · Assess Cisco MCP Scanner claude-code-security-review · AI SOC · Assess claude-code-security-review clawsec · Agent & MCP security · Assess clawsec codex-security · AI code security · Assess codex-security Counterfit · LLM red-teaming · Hold Counterfit cross-code-organizer · Agent & MCP security · Assess cross-code-organizer CubeSandbox · Agent & MCP security · Assess CubeSandbox Cybench · Benchmark · Trial Cybench E2B · Agent & MCP security · Assess E2B emilia-protocol · Agent & MCP security · Assess emilia-protocol Fickling · Model security · Trial Fickling garak · LLM red-teaming · Adopt garak Giskard · LLM red-teaming · Trial Giskard Guardrails AI · Guardrail · Trial Guardrails AI HexStrike AI · Offensive agent · Assess HexStrike AI hol-guard · Agent & MCP security · Assess hol-guard iFixAi · Benchmark · Assess iFixAi Inspect · Benchmark · Adopt Inspect Lakera Guard · Guardrail · Trial Lakera Guard LangKit · Guardrail · Hold LangKit lean-ctx · Agent & MCP security · Assess lean-ctx LLM Guard · Guardrail · Hold LLM Guard llm-attacks (GCG) · LLM red-teaming · Hold llm-attacks (GCG) lunar · Agent & MCP security · Assess lunar Mastyf · Agent & MCP security · Assess Mastyf mcp-context-protector · Agent & MCP security · Hold mcp-context-protector mcp-scan · Agent & MCP security · Trial mcp-scan MCPGuard-Dynamic · Agent & MCP security · Assess MCPGuard-Dynamic microsandbox · Agent & MCP security · Assess microsandbox Microsoft Security Copilot · AI SOC · Trial Microsoft Security Copilot ModelScan · Model security · Adopt ModelScan nasiko · Agent & MCP security · Assess nasiko Nebula · Offensive agent · Assess Nebula NeMo Guardrails · Guardrail · Adopt NeMo Guardrails nono · Agent & MCP security · Assess nono numbat · Agent & MCP security · Assess numbat onecli · Agent & MCP security · Assess onecli open-code-review · AI code security · Assess open-code-review OpenSandbox · Agent & MCP security · Assess OpenSandbox OpenShell · Agent & MCP security · Assess OpenShell opik · Model security · Assess opik Pensar Apex · Offensive agent · Assess Pensar Apex PentAGI · Offensive agent · Assess PentAGI PentestGPT · Offensive agent · Assess PentestGPT picklescan · Model security · Trial picklescan PINT Benchmark · Benchmark · Hold PINT Benchmark pipelock · Guardrail · Assess pipelock presidio · Guardrail · Assess presidio promptfoo · LLM red-teaming · Adopt promptfoo PurpleLlama (Llama Guard, Prompt Guard, LlamaFirewall) · Guardrail · Trial PurpleLlama (Llama Guard, Prompt Guard, LlamaFirewall) PyRIT · LLM red-teaming · Adopt PyRIT Rebuff · Guardrail · Hold Rebuff Red Clippy · Offensive agent · Assess Red Clippy sast-skills · AI code security · Assess sast-skills Shannon · Offensive agent · Assess Shannon ship-safe · AI code security · Assess ship-safe skill-scanner · Agent & MCP security · Assess skill-scanner skills · AI code security · Assess skills SkillSpector · Agent & MCP security · Assess SkillSpector skylos · AI code security · Assess skylos Strix · Offensive agent · Assess Strix substrate · Agent & MCP security · Assess substrate superagent · Guardrail · Assess superagent ThinkWatch · Guardrail · Assess ThinkWatch toolhive · Agent & MCP security · Assess toolhive vestige · Agent & MCP security · Assess vestige vibekit · Agent & MCP security · Assess vibekit Vigil · Guardrail · Hold Vigil visa-vulnerability-agentic-harness · AI code security · Assess visa-vulnerability-agentic-harness Weir · Agent & MCP security · Assess Weir Wirken · Agent & MCP security · Assess Wirken XBOW · Offensive agent · Assess XBOW
Offense Defense Dual-use Not yet reviewed Seen in attacks
Adopt
Mature and widely used; a sensible default for most teams.
Trial
Worth piloting on a real project; expect some rough edges.
Assess
Promising or important to understand; evaluate, don't depend on it yet.
Hold
Unmaintained, superseded or risky; avoid new adoption.

Reviewed tools

46
ToolFamilyRingSideStarsLast activityLicense
Adversarial Robustness Toolbox Adversarial Robustness Toolbox (ART) is a Python library from the Linux Foundation AI & Data project for evaluating and defending ML models against evasion, poisoning, extraction, and inference attacks. It supports major ML frameworks and data types, and is used by red and blue teams. LLM red-teaming Trial Dual-use 6,326 MIT
AgentDojo AgentDojo is a Python benchmark and dynamic environment for measuring prompt injection attacks and defenses against tool-using LLM agents across task suites such as workspace. It is used by researchers and developers to score agent utility and attack success under configurable defenses and attacks. Benchmark Trial Defense 1,065 MIT
Agentic Radar · stale Agentic Radar is an open-source CLI from SplxAI that statically scans agentic workflow code (LangGraph, CrewAI, n8n, OpenAI Agents, AutoGen) to map tools, MCP servers and workflow graphs to OWASP LLM and Agentic risks. Its test command runs simulated adversarial inputs against OpenAI Agents workflows. Agent & MCP security Hold Defense 1,069 Apache-2.0
AgentZ AgentZ is a multi-tenant platform for building and running AI agents, with default-deny Cilium network policies, a secret injection proxy that keeps credentials away from agents, and traceable agent actions. Aimed at teams deploying agents on Kubernetes. Agent & MCP security Assess Defense 56 Apache-2.0
ai-exploits · stale Protect AI's collection of Metasploit modules and Nuclei scanning templates for responsibly disclosed vulnerabilities in ML tooling (e.g., Ray, MLflow), plus CSRF templates. Intended for security professionals testing their own AI infrastructure. Model security Hold Offense 1,754 open source
ARTEX · withdrawn Seen in attacks · 12 mentions Multi-agent autonomous penetration-testing framework (Go, planner/worker design, works with hosted and local LLMs). Published openly in 2026, then abused by several China-linked groups in automated intrusions; the author withdrew the original code, but forks circulate. Offensive agent Hold Offense — — open source
Augustus Augustus is a Go-based LLM vulnerability scanner that runs adversarial probes (prompt injection, jailbreaks, encoding evasion, data extraction, agent attacks) against 28 LLM provider categories and produces reports for security teams. LLM red-teaming Assess Dual-use 321 Apache-2.0
BoxPwnr BoxPwnr is an open-source framework that runs LLMs or CLI agents (Claude Code, Codex, Grok, etc.) autonomously against CTF challenges and security labs across many platforms, and publishes solving traces and benchmark statistics. Benchmark Assess Dual-use 467 AGPL-3.0
CAI (Cybersecurity AI) · archived Open-source framework for building AI agents that automate offensive and defensive security tasks. The repository is archived: the final source tree is read-only, and no further releases or security patches are planned. Offensive agent Hold Dual-use 9,853 open source
Casco Commercial AI-driven penetration-testing platform whose agents test authorised scopes (web apps, APIs, cloud, mobile apps, MCP servers, AI apps and agents) and validate impact, with optional human review. Offensive agent Assess Dual-use — — commercial
Cisco MCP Scanner Cisco MCP Scanner is a Python CLI, REST server and SDK that scans MCP servers, tools, prompts and resources. It combines YARA rules, LLM-based analysis, Cisco AI Defense inspect API, VirusTotal hash lookups, and pip-audit/PyPI sandbox checks. Intended for MCP server developers and security teams. Agent & MCP security Assess Defense 1,117 Apache-2.0
Counterfit · archived Counterfit is an open-source automation layer from Microsoft Azure for assessing the security of machine learning systems. It unifies existing adversarial ML frameworks (text, tabular and image attacks) behind one CLI and Python interface, and lets users add their own targets and attacks. LLM red-teaming Hold Dual-use 941 MIT
Cybench Cybench is a benchmark and agent framework from Stanford and collaborators for evaluating model capabilities on 40 CTF-style cybersecurity tasks, with optional subtask-level scoring. It is used by researchers and AI labs to measure cyber-offense risk of LLM agents. Benchmark Trial Dual-use 355 Apache-2.0
Fickling · 1 mention Fickling is a decompiler, static analyzer, and bytecode rewriter for Python pickle files, including PyTorch model files. It detects malicious pickles, enforces import allowlists when loading ML models, and can also inject code into pickles for testing. Model security Trial Defense 677 LGPL-3.0
garak Open-source NVIDIA LLM vulnerability scanner that probes models and LLM-backed apps for prompt injection, data leakage, jailbreaks, hallucination and toxicity. Aimed at AI red teams and security assessors. LLM red-teaming Adopt Dual-use 10,128 Apache-2.0
Giskard Giskard is an open-source Python library for evaluating and red-teaming LLM and agentic systems. Giskard Checks provides evals with LLM-as-judge, and Giskard Scan runs automated adversarial vulnerability scans (prompt injection, jailbreaks, harmful content) against agents. LLM red-teaming Trial Defense 5,967 Apache-2.0
Guardrails AI Guardrails AI is a Python framework that wraps LLM calls in input and output guards built from validators (e.g., regex, toxicity, competitor checks) and can enforce structured output from LLMs. It can also run as a Flask/REST server. Guardrail Trial Defense 7,606 Apache-2.0
HexStrike AI HexStrike AI is an MCP server that exposes 150+ security tools (recon, web testing, password cracking, cloud audit, binary analysis) to AI agents such as Claude, Cursor, and VS Code Copilot, with autonomous agents for bug bounty, CTF, and exploit generation. Intended for authorised penetration testers and red teams. Offensive agent Assess Offense 13,482 MIT
Inspect Open-source framework from the UK AI Security Institute for evaluating LLMs and agents, with tool use, sandboxed agent tasks and model-graded scoring; widely used to run cyber-capability and safety evaluation suites. Benchmark Adopt Dual-use 3,212 MIT
Lakera Guard Commercial API that screens LLM inputs and outputs for prompt injection, jailbreaks, data leakage and harmful content before they reach the model or the user. Guardrail Trial Defense — — commercial
LangKit · stale LangKit is an open-source Python text-metrics toolkit that extracts signals from LLM prompts and responses (quality, relevance, jailbreak/prompt-injection/refusal similarity, toxicity, sentiment, regex patterns) and logs them via whylogs for monitoring, with optional WhyLabs platform integration. Guardrail Hold Defense 1,005 Apache-2.0
LLM Guard · archived Open-source Python toolkit by Protect AI with input and output scanners for LLM applications, covering prompt injection, sensitive data leakage, toxicity, banned topics, and malicious URLs. Used by developers integrating LLMs into production apps. The project is archived and no longer maintained. Guardrail Hold Defense 3,219 MIT
llm-attacks (GCG) · stale llm-attacks is the research codebase for the GCG (Greedy Coordinate Gradient) method, which generates adversarial suffixes that get aligned open-weight LLMs such as Vicuna and LLaMA-2-Chat to comply with harmful requests, with transfer to other models. It is aimed at ML researchers reproducing AdvBench experiments. LLM red-teaming Hold Offense 4,871 MIT
Mastyf Mastyf is a local reference monitor and security gateway for AI agents and MCP. It intercepts agent tool calls, applies pattern, schema, and optional LLM-based policy checks, and blocks or logs actions before they reach tool backends. Intended for teams running AI agents with MCP tools. Agent & MCP security Assess Defense 21 AGPL-3.0
mcp-context-protector · archived mcp-context-protector is an Apache-2.0 wrapper by Trail of Bits that sits in front of MCP servers to pin server configurations, block unapproved changes, scan tool responses with guardrail providers, and quarantine suspicious results. It targets users and teams running third-party MCP servers in MCP host apps. Agent & MCP security Hold Defense 228 Apache-2.0
mcp-scan Snyk Agent Scan discovers locally installed agent components (MCP servers, agent skills, and harnesses such as Claude Code, Cursor, and VS Code) and scans them for prompt injection, tool poisoning, and other risks. Intended for developers and security teams auditing agent configurations. Agent & MCP security Trial Defense 3,237 Apache-2.0
Microsoft Security Copilot Generative-AI assistant and agents for security operations: incident summarisation, guided investigation, threat-intel lookups and script analysis across Microsoft's security stack. AI SOC Trial Defense — — commercial
ModelScan ModelScan is an open-source Protect AI scanner that inspects ML model files (Pickle, H5, SavedModel, Keras and others) for unsafe embedded code without loading them. It is aimed at teams that consume third-party or pipeline-produced models. Model security Adopt Defense 793 Apache-2.0
Nebula Nebula is a cybersecurity harness for operators running authorized engagements. It bundles a project workspace, a security browser with interception, Kali tools in containers, and a Butler assistant that drives AI agents with scope and approval controls. Offensive agent Assess Offense 1,154 BSD-2-Clause
NeMo Guardrails NVIDIA NeMo Guardrails is an open-source Python toolkit for adding programmable input, dialog, retrieval, execution, and output rails to LLM-based conversational applications. It is aimed at developers building chatbots, RAG systems, and LLM endpoints. Guardrail Adopt Defense 7,484 open source
nono nono is an unprivileged, kernel-enforced sandbox and policy runtime for AI agents such as Claude Code, Codex and LangChain-based agents. It isolates tools, injects credentials via a proxy, filters network access, verifies skills with Sigstore, and keeps audit logs and file rollback. Agent & MCP security Assess Defense 4,911 Apache-2.0
Pensar Apex Pensar Apex is an open-source (Apache-2.0) terminal tool that runs autonomous AI agents to perform blackbox and whitebox penetration tests against target URLs, with headless CLI and CI/CD options. It is aimed at developers and security engineers. Offensive agent Assess Offense 324 Apache-2.0
PentAGI · 1 mention PentAGI is a self-hosted, open-source multi-agent system that autonomously plans and runs penetration-testing tools (nmap, Metasploit, sqlmap) inside a Docker sandbox, using a user-chosen LLM provider. It is aimed at security engineers and red teams testing systems they are authorised to assess. Offensive agent Assess Offense 33,803 MIT
PentestGPT PentestGPT is an open-source, LLM-driven penetration testing agent that runs multi-stage recon, exploitation and reporting pipelines via Claude Code or Codex, plus an interactive legacy mode across several LLM providers. Intended for CTFs and authorised pentests. Offensive agent Assess Offense 16,355 MIT
picklescan picklescan is an open-source CLI scanner that statically inspects Python pickle files (including PyTorch zip archives and remote Hugging Face or URL sources) for suspicious global imports such as eval or os calls. It is aimed at ML engineers and security teams vetting model files before loading them. Model security Trial Defense 429 MIT
PINT Benchmark · archived Lakera's PINT Benchmark is an evaluation dataset of about 4,314 English and non-English inputs for scoring prompt injection detectors such as guardrail APIs and classifier models. It supports custom datasets and evaluation functions. Benchmark Hold Defense 200 MIT
promptfoo · 1 mention Promptfoo is a CLI and library for running evals and red-teaming LLM-based apps and agents, including vulnerability scanning and reports. It is used by developers and security teams to test prompts, compare models, and gate CI/CD. LLM red-teaming Adopt Dual-use 27,536 MIT
PurpleLlama (Llama Guard, Prompt Guard, LlamaFirewall) PurpleLlama is Meta's open umbrella project providing LLM safeguard models (Llama Guard for input/output content moderation, Prompt Guard for prompt injection and jailbreak detection, Code Shield for inference-time insecure-code filtering) plus cyber-risk evals (CyberSec Eval 1-3). Aimed at developers building LLM applications. Guardrail Trial Defense 4,478 open source
PyRIT · 1 mention PyRIT is an open-source Python framework and toolkit from Microsoft for automated and multi-turn red teaming of generative AI models and applications. It offers a CLI scanner, a GUI, and extensible components for testing LLM endpoints, HTTP/WebSocket apps, and browser-based targets. LLM red-teaming Adopt Dual-use 5,021 MIT
Rebuff · archived Rebuff is an open-source prompt injection detector for AI applications that combines heuristics, an LLM-based classifier, a vector store of past attack embeddings, and canary-word leak detection. Intended for developers securing LLM apps; it also offers a hosted playground and a self-hostable server. Guardrail Hold Defense 1,532 Apache-2.0
Red Clippy Open-source pentest management server that an AI agent operates over MCP. It records scope, assets, observations, methodology coverage, and findings (with CVSS and evidence) in a database so engagement state persists across agent sessions. Offensive agent Assess Offense 30 GPL-3.0
Shannon Shannon is an open-source autonomous AI pentester for web apps and APIs. It combines source-code analysis with browser and CLI-driven exploitation of a running target and reports only findings with a working proof of concept. Intended for owners or explicitly authorised testers, including staging and release pipelines. Offensive agent Assess Offense 51,577 AGPL-3.0
Strix Strix is an open-source autonomous AI pentesting tool that runs multi-agent teams to reconnaissance, exploit, and validate web app, API, and code vulnerabilities via proof-of-concept exploits. It also offers auto-fix patches, reports, a local viewer, and CI integration. Offensive agent Assess Dual-use 74,691 Apache-2.0
Vigil · stale Vigil is a Python library and REST API that scans LLM prompts and responses for prompt injections, jailbreaks, and canary-token leakage using YARA rules, a transformer classifier, vector similarity, and sentiment analysis. Intended for developers self-hosting input/output filtering; it is alpha-stage and experimental. Guardrail Hold Defense 510 Apache-2.0
Weir Weir plants unique fake API keys into an AI agent's environment, runs the agent's flows against a local fake model, and checks every span, log, console output, model prompt and outbound request for those exact values. It also offers a Claude Code status line and redaction hook that track real keys entering or leaving sessions. Agent & MCP security Assess Defense 19 Apache-2.0
XBOW Commercial autonomous penetration-testing service: AI agents look for and validate web application vulnerabilities in scopes that customers authorise. Offensive agent Assess Dual-use — — commercial
Defensive tools found automatically, not yet reviewed (49)

Found from GitHub topics, curated awesome lists and news. Not rated by the editor yet; treat as untrusted code. Offensive and dual-use finds stay unpublished until reviewed.

ToolFamilyRingSideStarsLast activityLicense
open-code-review Open Code Review is a CLI that uses an LLM agent with tool access to review Git diffs and whole files, producing structured, line-positioned review comments. It targets developers and CI pipelines doing AI-assisted code review. AI code security Assess Defense 49,649 Apache-2.0
iFixAi iFixAi is an open-source Python auditor that runs simulated business scenarios against an AI agent's chat endpoint, sends it requests including out-of-role ones, and has a separate judge model grade each reply across 61 inspections. It is aimed at teams assessing agent alignment, access control and reliability before deployment. Benchmark Assess Defense 44,047 Apache-2.0
SkillSpector SkillSpector is a static and optional LLM-assisted security scanner for AI agent skills (SKILL.md bundles, repos, zips, MCP registry payloads). It reports vulnerabilities, malicious patterns and risk scores before a skill is installed. Intended for skill publishers, registries and install-gating pipelines. Agent & MCP security Assess Defense 24,890 Apache-2.0
opik Opik is an Apache-2.0 open-source platform for LLM and AI agent tracing, evaluation, prompt management, production monitoring, and guardrails. It is aimed at ML and engineering teams that want self-hostable observability for LLM apps. Model security Assess Defense 23,259 Apache-2.0
OpenShell OpenShell is a runtime that sandboxes autonomous AI agents. Operators declare file, network, and process rules in a policy, which is enforced at the kernel level, and credentials are injected only into requests to approved endpoints. Agent & MCP security Assess Defense 21,276 Apache-2.0
OpenSandbox Open-source sandbox runtime that runs AI agent code, commands, file operations, and browsers/desktops in isolated environments on Docker or Kubernetes. Adds per-sandbox egress policies and a credential vault for agent workloads. Agent & MCP security Assess Defense 16,651 Apache-2.0
E2B E2B is open-source infrastructure for running AI-generated code in isolated cloud sandboxes, controlled via JavaScript and Python SDKs. It is aimed at developers building AI agents and code-execution features, with self-hosting support on AWS, GCP, Azure, or generic Linux. Agent & MCP security Assess Defense 14,769 Apache-2.0
CubeSandbox CubeSandbox is an open-source, E2B-compatible sandbox service built on RustVMM and KVM that runs AI agent code in hardware-isolated MicroVMs with snapshot/rollback, egress filtering, and a credential-injecting security proxy. It is aimed at teams operating agent code-execution infrastructure. Agent & MCP security Assess Defense 13,193 open source
nasiko Nasiko is an open-source runtime and control plane for coding agents (Claude Code, Codex, OpenCode, Cursor CLI). It discovers installed agents, records their session turns and costs, and optionally routes their LLM calls through a configured provider. Agent & MCP security Assess Defense 11,944 open source
presidio Presidio is an open-source PII detection and de-identification SDK for text and images, usable from Python, PySpark, Docker, and Kubernetes. It is used by data teams to find and redact sensitive data in pipelines and in LLM inputs and outputs. Guardrail Assess Defense 11,779 MIT
codex-security · 1 mention Codex Security is an OpenAI CLI, TypeScript SDK, and GitHub Action that scans repositories for vulnerabilities using AI, validates findings, drafts patches and SECURITY.md policies, and exports results to trackers or SARIF. It is aimed at developers and security teams. AI code security Assess Defense 11,330 Apache-2.0
microsandbox Microsandbox is an open-source CLI and SDK that runs untrusted workloads, including AI agent code, in local microVMs with hardware isolation, OCI images, and network allow-listing. It is aimed at developers who need to sandbox agent-generated code. Agent & MCP security Assess Defense 8,988 Apache-2.0
superagent Superagent SDK is an open-source toolkit for AI agent safety. It provides runtime guard (prompt injection and unsafe tool-call detection), PII/secret redaction, repository scanning for agent-targeted attacks, and a planned red-team test feature for production agents. Guardrail Assess Defense 6,852 MIT
claude-code-security-review · stale A GitHub Action and /security-review slash command that uses Claude Code to semantically review pull request diffs for security vulnerabilities, posting findings as PR comments with false-positive filtering. Intended for maintainers of trusted repositories. AI SOC Assess Defense 6,802 MIT
AiSOC Open-source, self-hostable AI Security Operations Center that ingests telemetry, runs detection rules, groups alerts into incidents, and has AI agents investigate them with logged prompts and tool calls. Proposed actions need human approval. AI SOC Assess Defense 6,742 MIT
agent-governance-toolkit Agent Governance Toolkit is an open-source (MIT) policy enforcement, identity, audit, and sandboxing layer for autonomous AI agents. It intercepts tool calls, messages, and delegations via YAML/OPA/Cedar policies and provides Python, TypeScript, .NET, Rust, and Go SDKs plus CLI checks. Agent & MCP security Assess Defense 6,541 MIT
substrate Agent Substrate is a Kubernetes-based zero-trust runtime that sandboxes AI agents in gVisor or Cloud Hypervisor microVMs. It suspends idle agents, provides per-actor SPIFFE identity, default-deny networking, and an egress gateway that injects credentials outside the sandbox. Agent & MCP security Assess Defense 6,220 Apache-2.0
lean-ctx LeanCTX is an open-source context gateway that sits between AI coding tools (such as Claude Code, Cursor and Codex) and the files and tool output they read. It selects task-relevant context and applies configured path permissions, content filters and budgets, with local evidence records. Agent & MCP security Assess Defense 3,935 Apache-2.0
onecli OneCLI is an open-source platform for running sandboxed AI agents for teams. A gateway injects credentials into outbound requests so agents never hold secrets, and admins set one policy for all agents, with human approval for sensitive actions. Agent & MCP security Assess Defense 3,629 Apache-2.0
bla BLA is an open-source CLI and MCP server that hardens software build artifacts (JS, WASM, ASAR, SQLite, model files, and others) against AI reverse-engineering agents. It is aimed at software authors protecting their own releases. Agent & MCP security Assess Defense 3,580 MIT
visa-vulnerability-agentic-harness VVAH is an open-source Python pipeline that uses LLM agents to discover vulnerabilities in source code, adversarially verify each finding, propose or apply fixes, and have a review panel grade those fixes. It is aimed at AppSec teams triaging their own codebases. AI code security Assess Defense 3,059 open source
skill-scanner Cisco's Skill Scanner statically and behaviorally scans AI Agent Skills (Codex, Cursor, and lenient Claude Code formats) for prompt injection, data exfiltration, and malicious code, with an optional LLM judge. It is aimed at teams vetting agent skills before installing or committing them. Agent & MCP security Assess Defense 2,709 open source
toolhive ToolHive is an open-source platform for running MCP servers in isolated containers, with a gateway for per-request identity and access policy, a registry for curated servers, and a Kubernetes operator with audit logging and OpenTelemetry. It is aimed at developers, platform teams, and enterprises. Agent & MCP security Assess Defense 2,328 Apache-2.0
beelzebub Beelzebub is an open-source deception framework that deploys configurable decoy services (SSH, HTTP, TCP, TELNET, MCP) and captures attacker interactions as security evidence. It includes LLM-powered responses and MCP bait tools for observing suspicious agent tool invocations and prompt-injection attempts. Benchmark Assess Defense 2,232 GPL-3.0
agent-safehouse Agent Safehouse is a macOS sandboxing tool that wraps sandbox-exec with composable, deny-first policy profiles to restrict the files and integrations LLM coding agents can access. It targets developers running agents such as Claude Code, Codex, and Amp. Agent & MCP security Assess Defense 2,144 Apache-2.0
cc-safety-net CC Safety Net is a defensive hook for AI coding CLIs (Claude Code, Codex, Cursor, Gemini CLI and others). It parses shell commands and agent file-tool calls before they run, blocking destructive operations and access to secrets such as SSH keys and .env files. Agent & MCP security Assess Defense 1,976 MIT
vibekit · stale VibeKit is a CLI and SDK that runs coding agents such as Claude Code, Gemini CLI, and Codex inside isolated Docker sandboxes, with secret redaction and logging of agent operations. It is aimed at developers who run coding agents on their own machines. Agent & MCP security Assess Defense 1,868 MIT
akto Akto is a commercial platform for securing AI agents, MCP servers and LLM apps. Atlas covers employee endpoints (shadow AI discovery, guardrails, NHI governance); Argus covers homegrown AI (discovery, automated red teaming, runtime guardrails). Agent & MCP security Assess Defense 1,544 MIT
sast-skills A set of agent skills that let LLM coding assistants (Claude Code, Codex, Opencode, Cursor) run a static security review of a codebase. The project says it covers 13 vulnerability classes and writes a ranked report to sast/final-report.md. AI code security Assess Defense 1,364 MIT
skills · 2 mentions A set of AI agent skills for Claude Code, Cursor, Codex and similar tools that the Pashov Audit Group says cover Solidity security work: an x-ray pre-audit scan, a multi-agent Solidity auditor, and a fuzz-suite generator for Echidna and Medusa. AI code security Assess Defense 1,306 MIT
bernstein Bernstein is an Apache-2.0 governance layer for AI agents that enforces policy-as-code (who may do what, approval gates, required records) over CLI coding agents and other agent workloads. It uses a deterministic, model-free scheduler and an opt-in HMAC-chained audit log that supports offline verification. Agent & MCP security Assess Defense 1,197 Apache-2.0
Argus-Agentic-SOC-Platform Argus is an open-source, AI-native agentic SOC platform covering alert ingestion, tickets, CMDB, Sigma-based detection, workflow automation, and threat intel enrichment. Its AI Assistant exposes an MCP-style JSON-RPC tool registry for investigation tasks. AI SOC Assess Defense 1,178 Apache-2.0
numbat numbat is a Go CLI that monitors AI agent activity on endpoints via hooks, plugins, OTLP/HTTP logs, and on-disk session artifacts. It runs local CEL-based detection, offers opt-in pre-action blocking, and supports forensic reconstruction and case bundles. Agent & MCP security Assess Defense 1,163 Apache-2.0
clawsec ClawSec is an AGPL collection of security skills and signed advisory feeds for AI agent runtimes (OpenClaw, NanoClaw, Hermes, Picoclaw). It verifies skill artifacts, detects configuration drift, and approval-gates risky installs. Agent & MCP security Assess Defense 1,135 AGPL-3.0
hol-guard HOL Guard is an open-source, local-first security layer that reviews AI agent actions (shell commands, file access, package installs, MCP tool calls) and allows, blocks, or requests approval per policy. It targets developers running coding agents such as Codex, Claude Code, and Cursor. Agent & MCP security Assess Defense 1,078 Apache-2.0
pipelock Pipelock is an open-source AI agent firewall that sits between AI agents and the network, inspecting mediated HTTP, WebSocket, MCP, A2A and optionally TLS-intercepted traffic for secret exfiltration, prompt injection, SSRF, tool poisoning and risky tool-call chains. It also emits signed action receipts so reviewers can verify its decisions. Guardrail Assess Defense 987 Apache-2.0
skylos Skylos is an Apache-2.0 CLI and PR gate for Python (with partial multi-language support) that finds dead code, security bugs, secrets, vulnerable dependencies and AI-generated code mistakes. It also checks LLM app risks, and `skylos done` verifies whether an AI coding agent's branch is ready to merge. AI code security Assess Defense 888 Apache-2.0
ship-safe Ship Safe is a local CLI that scans repos for security issues, including AI/agent, MCP config, prompt, and RAG risks, then runs an investigation layer to classify findings as confirmed, likely, unresolved, or refuted. It also offers an interactive fix agent and CI gating with SARIF output. AI code security Assess Defense 870 MIT
ThinkWatch ThinkWatch is a self-hosted AI API and MCP gateway that authenticates model and MCP tool requests against an organization's identity provider, enforces rate limits and budgets, runs security guards (secrets redaction, dangerous-command checks, prompt-injection content filtering), and writes audit logs. Guardrail Assess Defense 815 open source
arcjet-js Arcjet JS SDK monorepo: runtime security checks called inside app code for HTTP request handlers (@arcjet/next, /node, /bun, etc.) and for non-HTTP paths such as AI agent tool calls and MCP handlers (@arcjet/guard). Covers prompt injection detection, sensitive information detection, rate limiting, bot protection, and custom rules. Guardrail Assess Defense 699 Apache-2.0
agent-safe-pipeline AgentSafe is an Apache-2.0 HTTP gateway that sits in front of AI agent and tool API calls, captures each consequential request as an intent, asks an external authorization control plane (Decionis) for ALLOW, ESCALATE or BLOCK, and forwards only authorized requests once. It includes a local demo authority and a request-testing command. Agent & MCP security Assess Defense 670 Apache-2.0
vestige Vestige is a local memory and guard layer for coding agents. It runs deterministic pre-execution rules on Bash, Edit and Write tool calls, logs failures and decisions to an append-only, hash-chained store, and provides causal lookup of past changes. Agent & MCP security Assess Defense 660 AGPL-3.0
Adrian Open-source (Apache-2.0) runtime monitoring engine for AI agents. It classifies tool calls, outputs, and reasoning traces against the agent's intended remit and can alert, hold for review, or block actions. Offers Python and TypeScript SDKs, a Claude Code plugin, and a self-hostable stack. Agent & MCP security Assess Defense 589 Apache-2.0
emilia-protocol Open-source authorization layer (EMILIA Gate, MCP guard, and scan tooling) that checks signed approval evidence and mandate limits before an AI agent's protected tool call reaches the provider. Aimed at teams gating high-impact agent actions such as payments, deployments, and repo deletion. Agent & MCP security Assess Defense 554 Apache-2.0
lunar Lunar.dev is an open-source API gateway and control layer for governing third-party API and LLM/agent traffic, including MCP server aggregation (Lunar MCPX), with policy enforcement, rate limiting, and cost/usage visibility. Intended for platform and security teams running AI agents. Agent & MCP security Assess Defense 509 MIT
agent-threat-rules ATR (Agent Threat Rules) is an open, YAML-based detection rule format and reference engine (TypeScript, with a Python wrapper) for threats against AI agents, covering prompt injection, tool poisoning, and credential exfiltration. It is aimed at teams building or operating agent runtimes, MCP integrations, and CI/SIEM pipelines. Agent & MCP security Assess Defense 426 MIT
cross-code-organizer Cross-Code Organizer (CCO) is a local dashboard for inventorying, editing, and cleaning up config for AI coding harnesses (Claude Code, Codex CLI, OpenCode, DeepSeek Harness). It includes an MCP security scanner that flags potentially poisoned tool descriptions and an experimental activation-probe scanner. Agent & MCP security Assess Defense 389 MIT
Wirken · 1 mention Wirken is an open-source gateway between messaging apps and AI agents. It provides identity management, per-channel isolation, a credential vault, and a per-session tamper-evident audit log, so a compromised agent has limited reach. Aimed at teams running autonomous agents in enterprise settings. Agent & MCP security Assess Defense 201 open source
MCPGuard-Dynamic · 1 mention A kernel-level eBPF sandbox that monitors and constrains the system activity of LLM agent tool calls made via the Model Context Protocol (MCP). Intended for teams running MCP tools in agent deployments. Agent & MCP security Assess Defense 75 open source

Following

Follow the vendors, products, AI tools, techniques and actors you care about. Matching stories, CVEs and exploited bugs are marked, a Following tab appears in the feed, and returning visits show what changed. Saved only in this browser.

Presets

Get updates

A personal RSS feed of new stories on everything you follow. Paste it into any feed reader, Slack or Teams RSS app.