agents-on-modal: Running Abliterated LLMs for Red Teaming

X postShrivu ShankarResource

Shrivu Shankar highlights abliterated open-weight models for autonomous security assessments and shares 'agents-on-modal'. The framework connects agent CLIs like Claude Code and Codex to self-hosted models running on Modal GPUs.

Key points

  1. Shrivu Shankar rates dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 as the top model for autonomous security testing.
  2. The open-source 'agents-on-modal' framework hooks Claude Code or Codex CLI into self-hosted models such as DeepSeek, GLM, and Qwen.
  3. An in-container gateway maps native Anthropic Messages or OpenAI Responses APIs directly to engine chat completions (SGLang/llama.cpp).
  4. Executions run across three sandbox types: local, isolated Docker containers, or massively parallel Modal cloud sandboxes.
  5. A 100-agent experiment successfully found IDOR flaws and credential misconfigurations across legacy web deployments.

Why it matters

Commercial LLM guardrails routinely block offensive security testing. Abliterated models overcome this for legitimate red teaming, provided they are safely contained.

Analysis

Background and Mechanism

Standard commercial foundation models reliably refuse red-teaming requests due to safety alignment. To conduct automated penetration testing, security practitioners increasingly turn to 'abliterated' models. By measuring refusal activations and subtracting that vector from model weights, models retain reasoning capabilities while removing guardrail refusals. Shrivu Shankar highlights dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 deployed on Modal inference (e.g. B300:4 GPUs).

Architecture of agents-on-modal

The companion repository sshh12/agents-on-modal decouples three operational components:

  • Models: Open-weight options (DeepSeek-V4, GLM-5.3, Qwen3.8) served via SGLang or llama.cpp on Modal GPUs with automatic scale-to-zero capabilities.
  • Harness: Compatibility with developer agent CLIs, specifically Claude Code and Codex CLI.
  • Sandbox: Runtime environments covering local execution, isolated throwaway Docker containers (full internet, no LAN access), or parallel Modal Sandboxes.

An in-container gateway bridges protocol differences, allowing harness tools to seamlessly interact with standard OpenAI-compatible endpoints. Sandboxes can also route egress traffic through residential HTTP/SOCKS5 proxies while keeping model token traffic direct.

Practical Findings

During an experiment tasking roughly 100 agent instances with hacking the author's accounts, agents uncovered IDOR vulnerabilities and credential issues in legacy hackathon projects and conducted OSINT and brute-force tests. The results highlight both the efficacy and attack surface introduced by autonomous agent workflows.

Original post

Shrivu Shankar

@ShrivuShankar · 7 November 2026

@dbmikus yes! dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 is the best i've tried use t.co/2s1QTpbPzl to run claude/codex on modal sandboxes on modal inference (B300:4) often red teaming my own apps on railway (pre-all the recent stuff thats happened w/them) t.co/86z5ZzArZX

66 likes3 replies157 bookmarks2.7k views

View on X
Selected replies (5)
  • @dbmikus Anybody here running abliterated models for security testing? I'm curious to learn about your setup: 1. which models? 2. which inference providers? 3. how do you secure the agent using the abliterated model? 4. how do you set up your pentest environment?
  • @refuselessai @dbmikus Might wanna try our abliterated GLM & DeepSeek
  • @NirvikBaruah @dbmikus Yup, been using abliterated deepseek in a sailbox for pentesting our infra. Model is self-hosted on our own GPUs and internal. I feed the networking logs + tool calls to a regular model so that it can periodically check what's going on and msg me summaries of what is happening.
  • @emil_no @dbmikus Tested both t.co/XHaCJoULVd and @OrcaCyberHQ
  • @anthonyronning @dbmikus dealign's dsv4.1 flash 1017 on @TryMapleAI on a kali box on an outside computer testing out dev infra need to work on better red team harnesses / goal / looping behavior because it's so quick, i have to keep responding to it's reports and say keep going lol.

Mentioned links

Links and tools from this entry

Summary generated by AI (Gemini 3.8 Flash, 8 November 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Repository:sshh12/agents-on-modal

    agents-on-modal: Run Coding Agents Against Open-Weight Models on Modal

    This Python utility connects coding agent CLIs like Claude Code and Codex to self-hosted open-weight LLMs on Modal serverless GPUs. It enables running isolated developer agents locally, in Docker, or via Modal sandboxes.

    5starsPython

    AI & LLMs· Tool

  • Link:Cantina Security

    Apex Flash 1 Abliterated: Open-Weights Security Model with Modified Refusals

    Cantina Security and Yeta Labs have released apex-flash-1-abliterated, an experimental open-weights model designed for authorised security research. Based on GLM-5.3-Flash, this multimodal variant alters refusal patterns across tasks. It assists researchers conducting audits in verified, permitted environments.

    AI & LLMs· Tool

  • Article:DeepInfra

    Top LLMs for OpenClaw Agentic Workloads

    DeepInfra evaluates open-weight models for the OpenClaw autonomous agent framework, focusing on tool-calling reliability, long-session instruction adherence, and cost per completed task.

    AI & LLMs· Resource

  • X post:Juraj Bednar

    Cloudflare Releases Multi-Phase Security Audit Skill for Coding Agents

    Cloudflare has open-sourced 'security-audit-skill', an agent framework that structures security audits through reconnaissance, adversarial validation, and verifiable reporting. Juraj Bednar highlights that using this structured skill enables even low-cost models to outperform unstructured prompts run on premium LLMs.

    16.4kbookmarks600.1kviews

    AI & LLMs· Tool

  • X post:Brais Moure

    Cloudflare Releases Open-Source Security Audit Skill for Coding Agents

    Brais Moure highlights the release of Cloudflare's 'security-audit-skill'. The tool turns coding agents into multi-phase security auditors and served as the foundation for Cloudflare's internal vulnerability discovery harness.

    3.4kbookmarks139.4kviews

    AI & LLMs· Tool

  • X post:O

    Post-Trained GLM-5.3-Flash Variants for Coding and Security

    In an online discussion on GLM-5.3-Flash, user O highlights community post-trained derivatives hosted on Hugging Face. UkisAI's Swift 1.5 significantly cuts reasoning tokens for general daily use, while Cantina Security's apex-flash-1-abliterated targets cyber research. The author ranks both variants ahead of stock GLM-5.3-Flash and DeepSeek.

    13bookmarks468views

    AI & LLMs· Discussion

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.