Open coreO·01
Strix (OmniSecure, Inc.) · github.com
Runs a team of AI agents that probe a codebase, live web app or API from a Docker sandbox, try to exploit what they find, and report validated findings with proof-of-concept exploits and suggested fixes.
- GitHub
- 74k stars · v1.7.0, 2026-10-05
- Licence
- Apache-2.0
- Tests
- Web apps, APIs, Code
- Runs as
- CLI with Docker sandbox (self-hosted); SaaS (Strix Cloud)
- Models
- OpenAI, Anthropic, Google (Gemini, Vertex AI), DeepSeek, Moonshot (Kimi), Z.ai (GLM)
- Price
- Pro $29/seat/month, pentests billed separately (pay per test); Enterprise custom pricing
Open coreO·02
Keygraph (Keygraph Inc.) · github.com
Reads a web application's source code to find likely attack paths, then attacks the running app and its APIs with browser automation and command-line tools, and reports only vulnerabilities it managed to exploit.
- GitHub
- 51k stars · v3.5.0, 2026-10-27
- Licence
- AGPL-3.0
- Tests
- Web apps, APIs, Code
- Runs as
- CLI (npx) with Docker worker (self-hosted); paid editions are SaaS or self-hosted
- Models
- Anthropic, OpenAI, xAI, AWS Bedrock, OpenAI-compatible gateways, local (Ollama, vLLM, LM Studio)
- Price
- Pro $50 per active developer/month; Enterprise custom
Open coreO·03
VXControl · github.com
A self-hosted multi-agent system: you describe a target and goal in plain language, and a team of AI agents plans the test, runs security tools inside an isolated Docker sandbox, keeps a memory of what worked, and writes a report.
- GitHub
- 34k stars · v2.2.0, 2026-10-05
- Licence
- MIT
- Tests
- Web apps, Networks
- Runs as
- Docker Compose (self-hosted web UI with REST and GraphQL APIs)
- Models
- OpenAI, Anthropic, Google (Gemini), AWS Bedrock, Azure OpenAI, DeepSeek
- Price
- Not published
Open sourceO·04
Gelei Deng (GreyDGL) and co-authors · github.com
An agent that drives Claude Code or Codex through staged reconnaissance, exploitation and reporting against a CTF box or pentest target; the original research version, which guides a human tester step by step, is kept as an interactive mode.
- GitHub
- 16k stars · v1.0.0, 2025-12-24
- Licence
- MIT
- Tests
- CTFs, Web apps, Networks
- Runs as
- CLI / Docker
- Models
- Anthropic (via Claude Code), OpenAI (via Codex), Google Gemini (interactive mode), DeepSeek (interactive mode), xAI (interactive mode), Qwen (interactive mode)
- Price
- Not published
Open sourceO·05
0x4m4 / OTT Cybersecurity LLC · github.com
An MCP server that lets an AI client such as Claude Desktop, Cursor or VS Code Copilot run a large catalogue of locally installed security tools against a target, with a decision engine that picks tools and parameters.
- GitHub
- 13k stars
- Licence
- MIT
- Tests
- Web apps, Networks, Cloud, CTFs
- Runs as
- MCP server (self-hosted Python)
- Models
- any MCP-compatible client (Claude Desktop, Cursor, VS Code Copilot, Roo Code, 5ire)
- Price
- Not published
Caution Check Point (2 Sep 2025) reported threat actors on dark-web forums claiming to use HexStrike AI to exploit three Citrix NetScaler flaws (CVE-2025-7775, CVE-2025-7776, CVE-2025-8424) within days of disclosure, cutting exploitation time 'from days to under 10 minutes'; The Hacker News reported the same on 3 Sep 2025. The README itself warns that it gives AI agents powerful system access and should run in isolated VMs. Source.
Open coreO·06
Alias Robotics · github.com
A Python framework for building AI agents that carry out offensive and defensive security work such as CTFs and bug-bounty testing; archived in 2026 as a research artifact.
- GitHub
- 9.9k stars
- Licence
- Mixed: MIT for code derived from openai-agents-python; Alias Robotics research-use licence (non-commercial research and academic use only) for its own additions
- Tests
- Web apps, APIs, Networks, CTFs
- Runs as
- CLI / Python library (pip install cai-framework)
- Models
- 300+ via LiteLLM, OpenAI, Anthropic, DeepSeek, local (Ollama)
- Price
- Not published
Caution The maker itself warns that the archived framework is unmaintained attack tooling: dependencies will age and the prompt-injection weaknesses it documented will not be fixed, so it should run only in isolated environments against authorised targets and never as part of a production security programme (its README). Source.
Open coreO·07
Bitter Security · github.com
An autonomous red-team agent that starts from an operator-approved scope and rules of engagement and runs multi-stage attack chains (recon, exploitation, privilege escalation, lateral movement, C2) from a Kali sandbox.
- GitHub
- 6.6k stars · v2.2.2, 2026-10-28
- Licence
- Apache-2.0
- Tests
- Web apps, Networks, Active Directory, Cloud
- Runs as
- Docker (CLI and web dashboard, self-hosted); SaaS (app.decepticon.red); Python SDK; MCP server
- Models
- Anthropic, OpenAI, Google Gemini, MiniMax, DeepSeek, xAI
- Price
- Solo $19/month; Pro $39/month (new Solo and Pro purchases currently closed; access by request)
Open sourceO·08
oritera (起零衍迹) · github.com
A general problem-solving engine in which worker agents share a graph of confirmed facts and planned next steps to search for a path from a starting point to a goal; penetration testing and CTFs are the first domain it was tested on.
- GitHub
- 4.9k stars · v0.2.1, 2026-05-10
- Licence
- AGPL-3.0
- Tests
- Networks, CTFs
- Runs as
- Docker Compose (server, dispatcher and worker containers); local mode without Docker
- Models
- Anthropic (via Claude Code), OpenAI (via Codex), Pi agent harness
- Price
- Not published
Open sourceO·09
Samuele Giampieri (samugit83) and contributors · github.com
A self-hosted framework that maps a target's attack surface into a Neo4j graph, lets an AI agent exploit it from a Kali sandbox behind human approval gates, then has a separate agent write fixes and open GitHub pull requests.
- GitHub
- 3.3k stars · v7.5.1, 2026-11-03
- Licence
- MIT
- Tests
- Web apps, APIs, Networks, LLM apps
- Runs as
- Docker Compose (self-hosted web app)
- Models
- OpenAI, Anthropic, OpenRouter, AWS Bedrock, any OpenAI-compatible endpoint (Ollama, vLLM, Groq)
- Price
- Not published
Open sourceO·10
GH05TCREW · github.com
A terminal-based agent framework for black-box testing that can follow single instructions, run a task on its own, or spawn a crew of worker agents, using a shell, a browser and prebuilt attack playbooks.
- GitHub
- 3.3k stars
- Licence
- MIT
- Tests
- Web apps, Networks
- Runs as
- CLI/TUI; Docker (base or Kali image)
- Models
- OpenAI, Anthropic, any LiteLLM provider, OpenAI-compatible relays
- Price
- Not published
Open sourceO·11
Armur AI · github.com
Runs many AI agents at once that coordinate through a shared findings board to discover, chain and verify vulnerabilities in web apps and APIs, then write a report.
- GitHub
- 3.2k stars · v0.3.7, 2026-10-30
- Licence
- AGPL-3.0
- Tests
- Web apps, APIs, Networks, CTFs
- Runs as
- CLI (npm, Homebrew, Go, Docker)
- Models
- Anthropic (Claude), OpenAI-compatible, Together AI, Google Gemini, local (Ollama, LM Studio)
- Price
- Not published
Open sourceO·12
zakirkun (Guardian Project) · github.com
A command-line tool in which planner, tool-selector, analyst and reporter agents choose and run conventional security tools against a target and turn their output into an evidence-linked report.
- GitHub
- 1.9k stars · v2.1, 2026-02-27
- Licence
- MIT, with added authorised-use terms
- Tests
- Web apps, APIs, Networks, Active Directory, Cloud, Mobile, Code, LLM apps
- Runs as
- CLI
- Models
- OpenAI, Anthropic, Google Gemini, OpenRouter, Requesty, local (Ollama)
- Price
- Not published
Open sourceO·13
0xSteph · github.com
Connects an AI client or model to a web-app and bug-bounty testing workflow, and only marks a finding verified when an independent check re-runs the exploit and compares it with a control.
- GitHub
- 1.8k stars · v1.6.1, 2026-11-03
- Licence
- MIT
- Tests
- Web apps
- Runs as
- MCP server and CLI (pipx install ptai); CI
- Models
- the model of your MCP client (Claude Code, Codex, Cursor, VS Code), local (Ollama), other providers per its provider guide
- Price
- Not published
Open coreO·14
BugBase · github.com
A browser-based agent that connects to a Kali attack box, runs tools, reads their output and decides the next step, aimed at real engagements, boot2root boxes and CTFs.
- GitHub
- 1.6k stars
- Licence
- MIT
- Tests
- Web apps, Networks, CTFs
- Runs as
- Docker (self-hosted web app); SaaS (Pentest Copilot Enterprise)
- Models
- OpenAI, Anthropic, Google, Mistral, Moonshot Kimi, any OpenAI-compatible endpoint
- Price
- Not published
Open sourceO·15
Joas A Santos · github.com
A Rust harness that runs a pool of LLMs through recon, exploitation and validation against a URL, source repo, host or AI endpoint, replays each claimed finding and has a second model check it before it reaches the report.
- GitHub
- 1.5k stars · v4.4.0, 2026-10-19
- Licence
- MIT
- Tests
- Web apps, Code, Networks, Active Directory, Cloud, LLM apps, Mobile
- Runs as
- CLI (plus TUI and web console)
- Models
- 19 providers via API key, Claude Code, Codex, Gemini and Grok subscriptions, local (Ollama, llama.cpp)
- Price
- Not published
Open sourceO·16
TU Wien IPA-Lab (Andreas Happe, Jürgen Cito and contributors) · github.com
A research framework for building small LLM-driven security-testing agents, shipping use cases for Linux and Windows privilege escalation, web and REST API testing, and Active Directory assumed-breach testing.
- GitHub
- 1.3k stars
- Licence
- MIT
- Tests
- Networks, Web apps, APIs, Active Directory
- Runs as
- CLI (wintermute) / Python library
- Models
- any LiteLLM provider, OpenRouter (default endpoint), OpenAI, Anthropic, Azure, local (Ollama)
- Price
- Not published
Open coreO·17
ASC-IT · github.com
A command-line tool that runs an autonomous assessment of an authorised target with LLM agents and records each finding with the exact command and raw output, optionally behind a local gateway that hides real IPs, hosts and credentials from the model.
- GitHub
- 1.2k stars · v1.6.0, 2026-10-25
- Licence
- GPL-3.0
- Tests
- Web apps, APIs, Networks, Active Directory, Cloud
- Runs as
- CLI with Docker (self-hosted); paid Pro web dashboard (cloud or self-hosted)
- Models
- Anthropic (Claude), OpenAI, OpenRouter, local (Ollama, llama.cpp)
- Price
- Pro EUR 149 per month (or EUR 1,788 billed annually); Pentest on Demand EUR 799 per engagement
Open sourceO·18
Beryllium Security · github.com
A desktop workspace for a human tester that keeps an engagement's assets, browser and intercepted traffic, Kali terminals, notes, evidence and reports in one project, with AI assistants and agents acting inside scope, approvals and budgets the operator sets.
- GitHub
- 1.1k stars · nebula-v3.0.45, 2026-11-08
- Licence
- BSD-2-Clause
- Tests
- Web apps, APIs, Networks
- Runs as
- desktop app (Linux x86_64 preview: DEB, AppImage, APT) plus headless core with a terminal client
- Models
- OpenAI (Codex harness; GPT-Live voice), xAI (Grok harness), Anthropic (Claude), OpenRouter, local (Ollama, vLLM)
- Price
- Not published
Open sourceO·19
Cyberspike (Changchun Anshanyuan Technology Co., Ltd.) · pypi.org
A Python package that researchers describe as an AI-native pentest framework acting as an MCP client, combining Kali Linux tools with DeepSeek models to scan, assess and generate exploits from natural-language commands.
- Licence
- None stated (PyPI metadata lists no licence)
- Tests
- Networks
- Runs as
- Python package (PyPI)
- Models
- DeepSeek
- Price
- Not published
Caution Straiker STAR Labs (11 Sep 2025) warned that Villager lowers the skill and time needed to run offensive toolchains, could follow Cobalt Strike's path from red-team tool to criminal and APT use, and uses containers set to self-destruct after 24 hours, wiping logs and evidence. The Hacker News (15 Sep 2025) reported the same concerns. Source.
Open coreO·20
Kindo (Usable Machines, Inc.) · deephat.ai
A family of open-weight language models fine-tuned for offensive and defensive cybersecurity work; it is a model you plug into an agent, not an agent that runs tests by itself.
- Licence
- Apache-2.0 plus a WhiteRabbitNeo/DeepHat usage-restriction extension (e.g. no military use); some versions use the Llama 3 or Llama 3.1 licence
- Tests
- Not stated
- Runs as
- model weights
- Price
- Not published