Open-source agent

SwarmAttacker vs CAI

An archived, research-licensed framework from Alias Robotics for building LLM security agents from a CLI, with a paid tier supplying its own low-refusal models.

Vendor
Alias Robotics
Licence
Dual: MIT for inherited code, research-use only for the rest
Pricing
Free research edition (bring your own key); CAI Pro at EUR 350 per month for commercial use.
Checked
September 2, 2026
CapabilitySwarmAttacker8 of 8CAI3 of 8
FreePartial.
Runs on a ChatGPT subscriptionNot found.
Custom prompts & skillsYes.
Safety-refusal handlingPartial.
Parallel multi-agent swarmYes.
MIT-licensed open sourceNot found.
Black-box from a URLPartial.
Model-agnosticYes.
  • 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 CAI's public pages or repository on September 2, 2026, not that it is technically impossible.

Profile

What is CAI?

CAI, or Cybersecurity AI, is a Python agent framework from Alias Robotics, a robot-cybersecurity company in Spain, first released in April 2025. It is built on a fork of OpenAI's agents library and organised around agents, tools, handoffs and patterns, shipping about thirty specialists across red, blue and purple teaming and offering parallel, swarm, hierarchical and sequential coordination patterns. Users drive it from an interactive command-line loop, switch agents and models at runtime, and can run unattended with a continue flag.

Its licensing is the catch: only the inherited OpenAI code is MIT, while everything Alias wrote is under a research-use licence that prohibits commercial, professional or production use without a paid licence, so despite a "MIT" classifier on some pages it is not open source in the usual sense. The free edition explicitly lists refusals from third-party models as a limitation and sells the fix, low-refusal proprietary models and abliteration-based steering, in the paid CAI Pro tier at EUR 350 per month. Model access is by API key through a broad provider layer; there is no ChatGPT-subscription login.

Alias reports strong capture-the-flag results and a substantial publication record. On 28 August 2026 the repository, with around 9,800 stars, was archived and squashed to a single commit with no further support, and work moved to a commercial successor. It is a general security-agent framework rather than a turnkey web pentester.

In SwarmAttacker's favour

Where it stands out.

  • 01A truly permissive MIT licence. CAI's own code is research-use only and bars commercial use without a paid licence, despite an "MIT" label on some pages.
  • 02Still maintained: SwarmAttacker is active, where CAI was archived in August 2026 and points users to a commercial successor.
  • 03Refusals are handled for free, where CAI lists them as a limitation and sells the fix in its paid tier.
  • 04Runs on a ChatGPT plan with no API key, which CAI does not support.
  • 05A single-purpose black-box web pentester, versus a general human-in-the-loop framework you assemble yourself.

In CAI's favour

Where it is stronger.

  • 01A general framework spanning capture-the-flag, operational technology, robotics and forensics, not only web apps.
  • 02A large publication record, documented prompt-injection defences and many capture-the-flag wins.
  • 03Access to more than 300 models and its own low-refusal models in the paid tier.
  • 04A mature community, with around 9,800 stars and many contributors, before archival.

Verdict

Which one should you pick?

Pick CAI if you want a general, research-oriented security-agent framework and can accept a research-use licence and an archived codebase. Pick SwarmAttacker if you need a maintained, MIT-licensed black-box web pentester that runs on a ChatGPT plan and handles refusals without a paid upgrade.

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
Source on GitHub