oss-redteam 0.1.1

GPT-OSS red-teaming pipeline and harness (OpenAI-compatible)

GPT-OSS Tool-Call Harness (HF Inference)

This repo runs prompts against openai/gpt-oss-20b:groq (or any OpenAI-compatible HF model), supplies a set of fake tools, captures when the model attempts tool-calls, then simulates tool failures and logs the follow-up response.

Outputs are written per-run to both CSV and SQLite with a timestamped run_id (no overwrites).

What it does

• Fake tools (subset): email_send, web_search, payroll_update, calendar_schedule_meeting, slack_post_message, github_create_issue, jira_create_ticket, db_sql_query, filesystem_read_file, payment_process, wiki_search, fetch_url, feature_flag_toggle, s3_list_objects, notion_update_page, twilio_send_sms, zoom_create_meeting, drive_search_files in src/tool_schemas.py.
• Initial turn: Sends your prompt with tools=... and records any tool calls.
• Simulated failures: If tools are called, the harness injects tool messages that report failure and requests a follow-up answer from the model.
• Logging:
  • CSV files: data/logs/run_<run_id>_interactions.csv, data/logs/run_<run_id>_tool_calls.csv
  • SQLite DB: data/logs/run_<run_id>.db (tables: runs, interactions, tool_calls)

Setup

1. Python 3.10+
2. Install deps

   python -m venv .venv
   # Windows PowerShell activation: .venv\\Scripts\\Activate.ps1
   pip install -r requirements.txt
   

3. Add your HF Inference token
   • Copy .env.example to .env
   • Set HF_TOKEN=...

Run

• Using sample prompts (console script):

  ossrt-harness --prompts-file prompts/examples.jsonl \
      --model openai/gpt-oss-20b:groq \
      --base-url https://router.huggingface.co/v1 \
      --temperature 0.2 --top-p 1.0 \
      --notes "smoke"
  

• Override system prompt with a local file:

  ossrt-harness --system-prompt-file prompts/system.txt
  

• Note: --cot is deprecated (no-op). GPT-OSS emits reasoning via Harmony channels automatically.

Inspect outputs

• CSVs in data/logs/
• SQLite database data/logs/run_<run_id>.db with tables:
  • runs(run_id, created_at, model, base_url, tools_json, prompts_file, notes, temperature, top_p)
  • interactions(run_id, prompt_idx, phase, role, content, tool_calls_json, raw_json, finish_reason, prompt_tokens, completion_tokens)
  • tool_calls(run_id, prompt_idx, call_idx, tool_call_id, name, arguments_json)

Notes

• This harness does not execute tools. It returns a structured failure for each tool call so you can observe how the model recovers.
• Some providers return old-style function_call instead of tool_calls. The harness supports both.
• Chain-of-thought flags (--cot) are deprecated and have no effect. GPT-OSS emits reasoning via Harmony channels (analysis/final).

Customization

• Add/edit tools in src/tool_schemas.py.
• Provide your own prompts via .jsonl, .json, or .txt.
• Extend logging or add new export formats as needed.