Skip to content

Back to the catalog

AI Security Briefing
github.comrepo

AI Security Briefing

HackAgent is an open-source Python SDK and CLI for testing the security of AI agents, automating research-backed attacks (e.g. jailbreaks, prompt injection) to surface vulnerabilities before exploitation. It runs locally with no API key required (SQLite-backed results and a built-in dashboard), supports an optional cloud reporting mode, and can drive browser-based agents via Playwright for WEB-provider targets. The project is Apache-2.0 licensed and framed explicitly for authorized security testing and coordinated disclosure.

TL;DR Open-source Python SDK/CLI that automates jailbreak and prompt-injection style attacks to red-team AI agents locally or via a cloud dashboard.
Open original ↗
Document

What it is

HackAgent, an Apache-2.0 licensed Python SDK and CLI for testing AI agents against agent-specific vulnerabilities. The summary covers the "what"; the page adds concrete setup mechanics not reflected there — particularly a non-obvious dependency step for browser-driven targets.

What to do

Standard install via pip in a virtualenv:

python3 -m venv .venv
source .venv/bin/activate
pip install hackagent

Alternatively, every GitHub Release ships a self-contained per-platform archive. Extract it and run the hackagent launcher inside — no Python, pip, or uv needed. Keep the extracted folder intact: the launcher loads the libraries sitting next to it, so moving the binary out of the folder breaks it.

The web dashboard is launched with:

hackagent web

Preconditions / Affected

The one real gotcha: the WEB provider drives a real browser through Playwright, and Playwright's browser binaries are not bundled in either the PyPI package or the release archives. Before testing any web-based target you must run once:

playwright install

Local mode requires no API key. Cloud sync to the hosted platform (https://app.hackagent.dev) is optional and only activates when an API key is configured.

Caveats

The project is explicitly scoped to authorized testing only — the page's usage policy requires explicit permission before testing any system, restricts use to your own agents or sanctioned pentests, and directs users to coordinated/responsible disclosure. Authors disclaim responsibility for misuse. Community contact is via GitHub discussions or ais@ai4i.it.

<system-reminder note omitted per instructions.>

Note: the "AI Security Briefing" title does not match the page, which is the HackAgent project repository — treat the catalog title with mild suspicion.

Technologies named here

Related in AI Security