Awesome Humanbound

Agent examples

Plain HTTP and FastAPI

• humanbound-quickstart FastAPI support bot that calls an OpenAI-compatible endpoint directly, with no agent framework in between. One file, uv run agent.py. Verified: D 43.14/100, 43 of 97 conversations failed, owasp_agentic, target gpt-4o-mini, 2026-09-04.

LangChain

• humanbound-langchain-example LangChain 1.x create_agent with order-lookup and refund tools behind a 15-line FastAPI wrapper. Talks itself into refunding orders that were never placed. Verified: F 27.38/100, 61 of 97 turns failed, restriction_bypass and human_manipulation dominant, owasp_agentic, 2026-09-10.

Wanted

Getting started

• Installation The extras decide what you get: engine for local testing, firewall for runtime defense.
• Quick start First run, end to end.
• Agent configuration The bot-config.json contract: chat_completion.{endpoint,headers,payload}, with $PROMPT and $CONVERSATION as the substitution tokens.
• Local engine Running without a Platform account, including against Ollama.
• Orchestrators What --quick, --deep, and --full actually select.
• LLM as judge How a conversation gets graded, and why one bad turn sinks the whole thing.
• Command reference Every hb subcommand.

Official

Repositories

• humanbound The adversarial testing engine, SDK, and hb CLI. Runs fully local or against the hosted Platform.
• humanbound-firewall Runtime firewall for agents. Cheap local tiers run first, and an LLM judge only gets called when they are uncertain.
• plugins Plugin marketplace for Claude Code and Cursor. It ships four agent skills, of which running-adversarial-tests is the one users trigger.
• actions Official GitHub Actions. Runs OWASP-aligned tests in CI and writes SARIF.

Packages and marketplace

• humanbound on PyPI pip install "humanbound[engine,firewall]". The CLI entry point is hb.
• humanbound-firewall on PyPI The firewall on its own, for runtime use without the test engine.
• Humanbound AI Agent Security Testing The Action on GitHub Marketplace. It writes a SARIF file but does not upload it, so pair it with a github/codeql-action/upload-sarif step.

Sites and services

• humanbound.ai Product site.
• docs.humanbound.ai Documentation, 44 pages covering the local engine, testing, defense, and integrations.
• Platform API reference Interactive OpenAPI docs for the hosted Platform.
• status.humanbound.ai Platform status page.
• llms.txt Machine-readable site index for agents and coding tools.

Self-assessment tools

• AI Security Readiness Checklist 23 questions, written for the person who has to sign off rather than the person who built the agent.
• Firewall Cost Calculator Estimates what the firewall tiers cost against your own traffic shape.

Writing

Tutorials

• Attack your own AI agent in under 10 minutes Ayan Pahwa, 2026-09-07. Builds a weak support agent, breaks it, patches it, and shows why the score barely moves afterwards.
• How to test a LangChain agent for security Ayan Pahwa, 2026-09-11. Wrapping an agent you already have in FastAPI so hb can reach it.

Essays

• Agent Security Debt: nobody is trying to break your AI agent until it ships Ayan Pahwa, 2026-08-31. Guardrails and governance and evals all exist. Adversarial testing before launch mostly does not, which is how IoT went wrong too.
• The enforcement illusion: why AI agent security starts with testing, not walls Kostas Siabanis, 2026-03-11. The argument behind testing before firewalling.
• Your agent passed its security test. That was three weeks ago. Kostas Siabanis, 2026-03-17. Why a point-in-time assessment goes stale faster for agents than for ordinary software.
• AI security means two different things Kostas Siabanis, 2026-04-27. Separates using AI for security from securing AI itself, which most vendor copy smears together.
• Why we open-sourced humanbound-firewall Demetris Gerogiannis, 2026-05-11. The reasoning, from the co-founder who made the call.
• Beyond moderation: why LLM systems need a policy layer Spyros Briakos, 2026-04-07. Where content moderation stops being enough for an agent that can act on its own.
• The agent attack scenario library Sofia Aliferi, 2026-08-20. The case for a shared, OWASP-mapped corpus of agent attacks.

Elsewhere

• The Taiwan attack: when an AI agent swarm ran a government hack with no one watching Sofia Aliferi, 2026-08-19.
• A new paper argues that your prompt injection defence can't win Sofia Aliferi, 2026-07-24.
• We put adversarial agent testing directly in Claude Code and Cursor Sofia Aliferi, 2026-07-23. The announcement behind the plugins repo.
• Trust Boundary Report, issue 02 Sofia Aliferi, 2026-08-04. A monthly roundup of agentic security incidents.

Talks

• Break first. Fix fast. Ship safe. Repeat. Demetris Gerogiannis at the Gen AI Unfold Summit, 2026-05-14. Continuous security testing inside the development cycle. Agenda listing only, no recording published.
• Securing agentic AI: before your agent betrays you Panel at AI and Beers Athens #6, 2026-05-28, with John Sotiropoulos, who leads the OWASP Agentic AI Top 10. No recording published.
• Ship agents that survive the real web Demetris Gerogiannis with Zyte, 2026-09-24. Upcoming at the time of writing. Zyte published slides and a recording for their previous meetup, so expect the same here.

Community

• Discord The project's community server.
• GitHub Discussions Questions and proposals against the engine.
• Contributing to Humanbound Upstream takes DCO sign-off (git commit -s) rather than a CLA.
• awesome-llm-security
• awesome-ai-security
• AI Red Teaming Guide

Related projects

• promptfoo LLM evaluation and red-teaming toolkit, with its own redteam_run.
• PyRIT Microsoft's Python Risk Identification Toolkit for generative AI.
• garak NVIDIA's LLM vulnerability scanner, organized around probes.
• agentic_security Agentic LLM vulnerability scanner and fuzzer.
• Scenario Simulation-based multi-turn agent testing from LangWatch.
• LLM Guard Runtime input and output guardrails, which puts it next door to humanbound-firewall.

Standards and regulation

• OWASP Top 10 for LLM Applications The taxonomy hb reports against.
• OWASP Agentic AI: threats and mitigations The agentic extension, which is what the owasp_agentic orchestrator targets.
• NIST AI Risk Management Framework Referenced by Humanbound's compliance mapping.
• EU AI Act explorer Searchable text of the regulation.