[GH-ISSUE #35] [Enhancement]: Is this pentagi is capable of automate internal penetration testing with adversary attacking plan #14

Open
opened 2026-06-06 22:07:58 -04:00 by yindo · 3 comments
Owner

Originally created by @peerapat-xfinit on GitHub (Apr 23, 2025).
Original GitHub issue: https://github.com/vxcontrol/pentagi/issues/35

Originally assigned to: @asdek on GitHub.

Target Component

Security Tools Integration

Enhancement Description

Hi' I'm curious about security feature of pentagi that capable of internal penetration test such Breach attack simulation such as CALDERA MITRE or something similar that can output as json or pdf report

Technical Details

No response

Designs and Mockups

No response

Alternative Solutions

No response

Verification

  • I have checked that this enhancement hasn't been already proposed
  • This enhancement aligns with PentAGI's goal of autonomous penetration testing
  • I have considered the security implications of this enhancement
  • I have provided clear use cases and benefits
Originally created by @peerapat-xfinit on GitHub (Apr 23, 2025). Original GitHub issue: https://github.com/vxcontrol/pentagi/issues/35 Originally assigned to: @asdek on GitHub. ### Target Component Security Tools Integration ### Enhancement Description Hi' I'm curious about security feature of pentagi that capable of internal penetration test such Breach attack simulation such as CALDERA MITRE or something similar that can output as json or pdf report ### Technical Details _No response_ ### Designs and Mockups _No response_ ### Alternative Solutions _No response_ ### Verification - [ ] I have checked that this enhancement hasn't been already proposed - [x] This enhancement aligns with PentAGI's goal of autonomous penetration testing - [x] I have considered the security implications of this enhancement - [ ] I have provided clear use cases and benefits
yindo added the enhancement label 2026-06-06 22:07:58 -04:00
Author
Owner

@asdek commented on GitHub (Apr 23, 2025):

Hello @peerapat-xfinit

Thank you for your interest in PentAGI 😉

I see that your inquiry touches on two different points, so I'll address them separately:


1. Output as JSON or PDF Report:

Regarding the capability to output reports in JSON or PDF format, there is an existing Issue (Issue #6) that tracks this reporting functionality. I'm keeping this in mind and am consistently making changes to the product to implement this feature.

In the next few days, I'm rolling out a significant update that introduces assistant functionalities to PentAGI. With this update, not all tasks need to be entirely delegated to agents—you'll have the ability to precisely control and intervene in the attack simulation process. Think of it as a "professional mode" that allows for more granular control. I believe this will not only enhance usability for advanced users but also help reduce expenses on LLM providers by preventing agents from performing unnecessary work.

This update brings substantial changes under the hood. I'm introducing mechanisms for summarizing message chains and refining the logic of agent interactions. In my testing, these improvements have already resulted in a 60% reduction in automation costs. Additionally, I've introduced a "simulation context" object that consolidates all executed steps, which greatly assists in performing subsequent subtasks.

Such foundational changes are essential before I expand functionality further, including reporting features and enhanced data visualization. Building a robust core ensures that as I add more features, I won't face structural issues that are difficult to adjust later.

There's also another fundamental change in agent logic to support reporting: memory management through the UI. This will allow you to conveniently access and edit lists of all artifacts collected during the attack simulation process.


2. Internal Penetration Testing like Breach Attack Simulation (e.g., CALDERA MITRE):

Your second question is a bit more complex. Traditional Breach and Attack Simulation (BAS) solutions, like CALDERA MITRE, are based on heuristics and predefined attack scripts. I'm currently exploring the idea of empowering agents to create their own attack scripts and execute them in an automated manner. However, at this stage, these are still conceptual ideas.

If you have specific suggestions or preferences on how you'd like this capability to function, I'd love to hear your thoughts. Your feedback can be incredibly valuable, and I'll do my best to incorporate it into the design and development process.

<!-- gh-comment-id:2823692595 --> @asdek commented on GitHub (Apr 23, 2025): Hello @peerapat-xfinit Thank you for your interest in PentAGI 😉 I see that your inquiry touches on two different points, so I'll address them separately: --- **1. Output as JSON or PDF Report:** Regarding the capability to output reports in JSON or PDF format, there is an existing Issue ([Issue #6](https://github.com/vxcontrol/pentagi/issues/6)) that tracks this reporting functionality. I'm keeping this in mind and am consistently making changes to the product to implement this feature. In the next few days, I'm rolling out a significant update that introduces assistant functionalities to PentAGI. With this update, not all tasks need to be entirely delegated to agents—you'll have the ability to precisely control and intervene in the attack simulation process. Think of it as a "professional mode" that allows for more granular control. I believe this will not only enhance usability for advanced users but also help reduce expenses on LLM providers by preventing agents from performing unnecessary work. This update brings substantial changes under the hood. I'm introducing mechanisms for summarizing message chains and refining the logic of agent interactions. In my testing, these improvements have already resulted in a 60% reduction in automation costs. Additionally, I've introduced a "simulation context" object that consolidates all executed steps, which greatly assists in performing subsequent subtasks. Such foundational changes are essential before I expand functionality further, including reporting features and enhanced data visualization. Building a robust core ensures that as I add more features, I won't face structural issues that are difficult to adjust later. There's also another fundamental change in agent logic to support reporting: memory management through the UI. This will allow you to conveniently access and edit lists of all artifacts collected during the attack simulation process. --- **2. Internal Penetration Testing like Breach Attack Simulation (e.g., CALDERA MITRE):** Your second question is a bit more complex. Traditional Breach and Attack Simulation (BAS) solutions, like CALDERA MITRE, are based on heuristics and predefined attack scripts. I'm currently exploring the idea of empowering agents to create their own attack scripts and execute them in an automated manner. However, at this stage, these are still conceptual ideas. If you have specific suggestions or preferences on how you'd like this capability to function, I'd love to hear your thoughts. Your feedback can be incredibly valuable, and I'll do my best to incorporate it into the design and development process.
Author
Owner

@ozgurozkan123 commented on GitHub (Feb 24, 2026):

@asdek If you want pentester agent to come up with novel attacks you need abliterated model like Pingu Unchained 3 or 4. Check these out it's available via API as well

one of the best models out there specifically fine tuned for this task is Pingu Unchained
https://penclaw.ai it uses Pingu Unchained 4 LLM
https://platform.audn.ai serves the Pingu Unchained 4 LLM
https://pingu.audn.ai to try out commands to Pingu Unchained 3

This is closed source model, the reason behind it there are further advancements to open source models such as internal tunning, specific GPU inference, custom attention implementation and internal data.

I am fan of open source but there's a reason people are using Claude Opus 4.6 ( a closed source model ) for multiple reasons.

<!-- gh-comment-id:3951098208 --> @ozgurozkan123 commented on GitHub (Feb 24, 2026): @asdek If you want pentester agent to come up with novel attacks you need abliterated model like Pingu Unchained 3 or 4. Check these out it's available via API as well one of the best models out there specifically fine tuned for this task is Pingu Unchained https://penclaw.ai it uses Pingu Unchained 4 LLM https://platform.audn.ai serves the Pingu Unchained 4 LLM https://pingu.audn.ai to try out commands to Pingu Unchained 3 This is closed source model, the reason behind it there are further advancements to open source models such as internal tunning, specific GPU inference, custom attention implementation and internal data. I am fan of open source but there's a reason people are using Claude Opus 4.6 ( a closed source model ) for multiple reasons.
Author
Owner

@asdek commented on GitHub (Mar 19, 2026):

hey @ozgurozkan123

thanks for the recommendation!

we've conducted internal testing of about a dozen different open source models and settled on qwen3.5-27b-fp8 for several reasons:

  • fully open source with permissive licensing
  • no strict cybersecurity usage restrictions
  • excellent performance on xbow benchmarks (results coming soon)
  • cost-effective inference with vLLM

we've already prepared a comprehensive guide on running PentAGI with this model via vLLM:
https://github.com/vxcontrol/pentagi/blob/master/examples/guides/vllm-qwen35-27b-fp8.md

general custom provider configuration is documented here:
https://github.com/vxcontrol/pentagi?tab=readme-ov-file#custom-llm-provider-configuration

while proprietary models like Pingu Unchained or Claude Opus 4.6 may offer certain advantages, for our use case the combination of open licensing, strong performance, and deployment flexibility made qwen3.5-27b-fp8 the right choice.

that said, PentAGI supports any OpenAI-compatible endpoint, so users are free to integrate Pingu Unchained or other specialized models if they prefer.

<!-- gh-comment-id:4094581178 --> @asdek commented on GitHub (Mar 19, 2026): hey @ozgurozkan123 thanks for the recommendation! we've conducted internal testing of about a dozen different open source models and settled on **qwen3.5-27b-fp8** for several reasons: - fully open source with permissive licensing - no strict cybersecurity usage restrictions - excellent performance on xbow benchmarks (results coming soon) - cost-effective inference with vLLM we've already prepared a comprehensive guide on running PentAGI with this model via vLLM: https://github.com/vxcontrol/pentagi/blob/master/examples/guides/vllm-qwen35-27b-fp8.md general custom provider configuration is documented here: https://github.com/vxcontrol/pentagi?tab=readme-ov-file#custom-llm-provider-configuration while proprietary models like Pingu Unchained or Claude Opus 4.6 may offer certain advantages, for our use case the combination of open licensing, strong performance, and deployment flexibility made qwen3.5-27b-fp8 the right choice. that said, PentAGI supports any OpenAI-compatible endpoint, so users are free to integrate Pingu Unchained or other specialized models if they prefer.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: vxcontrol/pentagi#14