Academic system
SwarmAttacker vs MAPTA
A University College London research system in which a GPT-5 coordinator delegates to sandboxed agents and a validation agent to produce verified proof-of-concept exploits.
- Vendor
- David & Gervais, University College London
- Licence
- MIT (partial, prompt-redacted code release)
- Pricing
- Free code; requires a paid OpenAI API key.
- Checked
- September 2, 2026
| Capability |  SwarmAttacker8 of 8 | MAPTA2 of 8 |
|---|---|---|
| Free | Yes. | |
| Runs on a ChatGPT subscription | Not found. | |
| Custom prompts & skills | Partial. | |
| Safety-refusal handling | Not found. | |
| Parallel multi-agent swarm | Partial. | |
| MIT-licensed open source | Partial. | |
| Black-box from a URL | Yes. | |
| Model-agnostic | Not found. |
- yes
- not found or only partly
Hover a mark for the evidence. A cross means the capability was not found, or only partly found, on MAPTA's public pages or repository on September 2, 2026, not that it is technically impossible.
Profile
What is MAPTA?
MAPTA, short for Multi-Agent Penetration Testing AI for the Web, is a research system from Isaac David and Arthur Gervais at University College London, published in August 2025. It describes a three-role architecture: a coordinator for strategy and delegation, one or more sandbox agents executing commands inside a per-job Docker container, and a validation agent that reproduces candidate findings into pass-or-fail proof-of-concept exploits. The system is driven exclusively by GPT-5 at high reasoning effort through the OpenAI API, so it needs an API key and offers no other model.
The paper's black-box evaluation took only a URL and a challenge description per target, and it states that those runs used a single coordinator agent; the full multi-agent pipeline was reserved for a separate white-box study of open-source applications. The paper does not discuss model safety refusals.
The MIT-licensed repository has two commits, no README, and a main system prompt that ships as a redacted placeholder, with a missing sandbox factory and several open issues asking for reproduction details that have gone unanswered. Its value is as a cited academic baseline with full cost and token accounting and real-world CVE findings, rather than a runnable tool.
In SwarmAttacker's favour
Where it stands out.
- 01A runnable, maintained tool. MAPTA's repo is two commits with a redacted prompt and unanswered issues.
- 02Model choice and a ChatGPT-subscription option, where MAPTA hardcodes the OpenAI API.
- 03Built-in refusal handling, which MAPTA does not address.
- 04A real parallel swarm on every run, where MAPTA's black-box evaluation used a single coordinator agent.
In MAPTA's favour
Where it is stronger.
- 01A peer-cited paper with full cost and token accounting and real-world CVE findings.
- 02A validation agent that reproduces every finding into a pass-or-fail exploit before reporting.
- 03A clear academic contribution that later systems use as a baseline.
Verdict
Which one should you pick?
MAPTA is a research reference, not a tool you would deploy: cite it for its methodology and cost figures. Pick SwarmAttacker if you want a maintained, higher-scoring black-box agent you can actually run, on your choice of model, with refusals handled.
Sources
What this page was checked against.
All checked on September 2, 2026.
Turn a URL into a security report.
Open source, MIT licensed, and it runs on the ChatGPT plan you already pay for. Install it and point it at a target you own.
brew install joloooo/swarm/swarm