H3llKa1ser

Phish Triage Agent — Development skill for Claude Code

Development community

Agentic triage for user-reported phishing emails.

How to install Phish Triage Agent

This entry records only its repository, not the path inside it, so there is no exact command to give. Open H3llKa1ser/phish-triage-agent and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Phish Triage Agent does

Agentic triage for user-reported phishing emails. Drop in an .eml, and a pair of Claude agents investigate it, enrich its indicators, write an analyst summary with their own verdict, and hand the case to a human for the final call. Nothing is ever actioned automatically.

Alternatives in Development

Investigate — Investigation Skill 1.9k ★
Investigate Performance — Investigate a Nimbalyst performance problem (freeze, lag, idle CPU, slow op, memory growth) with measured evid 1.6k ★
Pool Compare — Compare all Meteora DLMM pools for a token pair by APR, fee/TVL ratio, and volume 722 ★

README

phish-triage-agent

Agentic triage for user-reported phishing emails. Drop in an `.eml`, and a pair of Claude agents investigate it, enrich its indicators, write an analyst summary with their own verdict, and hand the case to a human for the final call. Nothing is ever actioned automatically.

flowchart LR
    A[.eml] --> B[Parser
deterministic, no LLM]
    B --> C[Investigator agent
Claude + read-only tools]
    C -->|lookup_url / lookup_domain / lookup_file_hash| T[(VirusTotal, RDAP)]
    C --> D[Coverage check
enrich anything skipped]
    D --> E[Critic agent
independent second opinion]
    E --> F[Policy guardrails
floors, hallucination check, priority]
    F --> G[Report
defanged Markdown]
    G --> H{{Human review queue
file or Jira}}
    H -->|approve / override| I[(Decision log
agent vs human stats)]

Stages

Stage	What it does	LLM?
Parser (`parser.py`)	Headers, SPF/DKIM/DMARC, Received IPs, URLs (text, HTML, and inside HTML/SVG attachments), attachment hashes, and 15+ heuristics: lookalike domains, link-text mismatch, reply-to mismatch, credential forms, risky extensions, zero-width chars, prompt-injection text.	No
Investigator (`agents.py`)	Tool-using agent. Decides which indicators to enrich, then submits a structured verdict (`benign / spam / suspicious / malicious`, confidence, summary, findings, MITRE, recommended actions) via a `submit_verdict` tool.	Yes
Coverage check (`policy.py`)	Enriches every attachment hash and the sender domain if the agent skipped them.	No
Critic (`agents.py`)	A second agent checks every finding against the evidence and can disagree. It cannot change the verdict; disagreement flags the case.	Yes
Policy (`policy.py`)	Raises the verdict when hard evidence demands it (known-bad hash, bad URL reputation, 2+ high-severity findings, injection attempt). Flags low confidence, critic disagreement, and domain