macOS
brew install humanboundlocal Homebrew formula metadata
brew / rank 11927
Adversarial security testing engine, SDK, and CLI for AI agents. Version 2.8.0 via Homebrew; verified 2026-07-31.
install
brew install humanboundlocal Homebrew formula metadata
overview
Adversarial security testing engine, SDK, and CLI for AI agents
history
Humanbound is an open-source adversarial testing engine, Python SDK, and hb command-line interface for AI agents. It drives multi-turn attacks against live agent endpoints, evaluates failures, and can turn findings into runtime guardrail rules.
The official Humanbound repository was created in February 2026 and is maintained by the verified Humanbound GitHub organization. The project combines an Apache-licensed local engine with an optional hosted platform, allowing initial testing without a login while preserving a path to managed posture tracking.
The official release notes describe a consolidation around the humanbound PyPI package: the earlier humanbound-cli package was yanked, and the CLI and public SDK now ship together. The same repository also integrates with the sibling humanbound-firewall package for a test-to-guardrail workflow.
Humanbound is a young project, so its reliable adoption history is limited. Official materials show distribution through PyPI, Homebrew in the supplied package facts, and an official GitHub Action for security gates in continuous integration.
Users describe an agent endpoint in a JSON configuration file and optionally define permitted and restricted behavior in a scope file. The hb test command runs attacks locally or through the platform; hb posture, hb logs, and hb report inspect the results.
Local attack generation and judging can use OpenAI, Anthropic, Gemini, Azure OpenAI, Grok, or Ollama. Provider settings resolve from CLI flags, environment variables, and then ~/.humanbound/config.yaml; platform login credentials are stored under ~/.humanbound/.
Humanbound is notable as an AI-security tool that packages an engine, SDK, and CLI under one Python distribution while keeping the firewall as a composable sibling package. Its optional extras let package users choose the provider engine, firewall runtime, or both.
security posture
No matching local secret-handling manifest was found for humanbound. Nucleus package metadata is still published here so future coverage has a stable package URL.
Before unattended agent use, check whether the tool reads plaintext credentials, writes remote state, publishes artifacts, or shells out to plugins.
local files
These source-backed paths show where this package keeps local settings or durable credentials. Automic Vault can use them as review targets for secret scanning, migration, and command approval.
Config paths the tool may read or write during local use.
~/.humanbound/config.yamlCredential-bearing paths to review before unattended agent runs.
~/.humanbound/executables
| Command | Kind | Exposure | Note |
|---|---|---|---|
hb | executable | indexed executable | Discovered from the local executable index. |
freshness
These signals separate page generation age, package-manager activity, and upstream release comparison. Version lag is warned only when an evidence URL and comparable versions are present.
install metadata
| Package key | brew:humanbound |
|---|---|
| Version | 2.8.0 |
| Package manager | Homebrew |
| Homepage | https://docs.humanbound.ai/ |
| Last updated | 2026-07-31T00:17:01Z |
| Pulse | updated |
| Bottle | not recorded |
| Service | none declared |
source trail
This page is generated by av-web from the private package SQLite artifact built by scripts/generate-pkg-sqlite.py.
View the package source record on GitHub.