A professional phishing detection agent stack based on OpenAI Agents SDK.

• Multi-agent workflow:
  • Router Agent: select FAST | STANDARD | DEEP
  • Investigator Agent: deep artifact analysis (text/URL/domain/attachment)
  • Summarizer Agent: final verdict + risk score + actions
• End-to-end attack-chain analysis:
  • email text + html parsing
  • URL safe fetch (sandbox policy + redirect chain + html signals)
  • attachment deep analyzer (PDF/image/audio/office/html static-safe pipeline)
  • domain intelligence (typosquat/punycode/risky-tld heuristics)
  • deterministic risk fusion (text + url + domain + attachment + ocr)
• Extensible tool architecture:
  • built-in tools
• Model access strategy:
  • openai: native OpenAI Agents SDK model path
  • local (and non-OpenAI providers): unified through LiteLLM
  • local runtime: Ollama

uv run python -m phish_email_detection_agent

Install OCR/audio analysis dependencies (optional):

uv sync --extra analysis

Single input text:

uv run python -m phish_email_detection_agent --text "Please verify your account now"

Structured deep input (text + urls + attachments):

uv run python -m phish_email_detection_agent --text '{"text":"Urgent: login now","urls":["https://bit.ly/reset"],"attachments":["invoice.zip"]}'

EML input:

uv run python -m phish_email_detection_agent --text '{"eml_path":"/path/to/sample.eml"}'

API safety note:

• eml_path is supported for local CLI usage, but rejected by API /analyze.
• API JSON attachments must be structured objects with name or filename.
• Path-like attachment values are rejected by API input validation.
• API supports debug_evidence=true to return full unsanitized evidence details; default responses are sanitized.

Argis now uses the same lightweight convention as SkillsBench: each skillpack is a folder containing SKILL.md under local skillpacks/.

List installable remote skillpacks from SkillsBench:

python scripts/skillsbench_skillpacks.py --list

Install specific skillpacks into project skillpacks/:

python scripts/skillsbench_skillpacks.py --install threat-detection openai-vision image-ocr

By default runtime auto-discovers local skillpacks from skillpacks/. You can override path with:

export MY_AGENT_APP_SKILLPACKS_DIR=/path/to/skillpacks

API /analyze responses now include both runtime.installed_skillpacks and top-level skillpacks summary (dir, count, names, installed). It also includes runtime.builtin_tools and top-level tools summary (count, names, builtin).

Safe defaults: URL fetch is disabled, private-network access is blocked, OCR/audio transcription are off.

# One switch: enable full deep analysis pipeline with built-in defaults
export MY_AGENT_APP_ENABLE_DEEP_ANALYSIS=true

If you need fine-grained control later, you can still override individual options (backend/model/limits) via env vars.

Default runtime is local Ollama (profile=ollama, model=ollama/qwen2.5:7b). So without extra env vars, the app prefers local inference over OpenAI API.

OpenAI:

export MY_AGENT_APP_PROFILE=openai
export OPENAI_API_KEY=your_key
uv run python -m phish_email_detection_agent --text "review this email"

LiteLLM + Ollama (local):

ollama pull qwen2.5:7b
ollama pull llama3.1:8b
uv run python -m phish_email_detection_agent --text "review this email"

Temporary model override:

export MY_AGENT_APP_PROFILE=ollama
uv run python -m phish_email_detection_agent --model ollama/qwen2.5:3b --text "review this email"

Quick local verification:

ruff check src tests docs scripts
pytest -k 'not hf_phishing_email_balanced_sample'