Phishing Investigator Autogen
blueprint · 6 agents · 4 MCP servers
964★ stars
124forks
4.7rating
What it does
Three analysts run in parallel on a reported email: a header analyst reads the auth and routing, a link analyst opens the URLs in a sandboxed browser and checks their reputation, and an attachment analyst scans payloads against threat intelligence. A prosecutor and a defender then debate the evidence to a verdict, which suppresses false alarms. A confirmed-malicious verdict is packaged with its evidence and paused for a human to approve remediation. Nothing is quarantined without sign-off.
The cast · 6 agents
Defenderdefenderreasoning
Argue that the email is benign, challenging weak or circumstantial evidence. Return an object with a 'malicious' boolean and a confidence level representing where the debate settled. Default to malicious only when the evidence is clear; flag genuine uncertainty for a human.
no tools · reasoning only
Packagerpackagerfast
Package the verdict and the evidence into a short case for the security channel. If malicious, list the recommended remediation (quarantine, block sender, reset if credentials were entered). Do not execute any remediation.
chat.post_message
Prosecutorprosecutorreasoning
Argue that the email is malicious, citing the strongest evidence from the three analysts. If you received a defender rebuttal, sharpen or concede specific points honestly.
no tools · reasoning only
Link AnalystlinkAnalystreasoning
Open each link in the sandboxed browser and check its reputation. Report the final landing domain, any credential-harvest form, redirects, and the reputation verdict. Never enter credentials or submit forms.
browser.openintel.url_report
Header AnalystheaderAnalystfast
Read the email headers. Report SPF, DKIM, and DMARC results, the true sending path, and any display-name or reply-to mismatch. Facts only.
mail.get_message
Attachment AnalystattachmentAnalystfast
Scan each attachment against threat intelligence. Report file type, hash reputation, and any macro or executable content. Do not execute attachments.
intel.file_report
Flow
⇉ parallel
→ Header Analyst → headerFindings
→ Link Analyst → linkFindings
→ Attachment Analyst → fileFindings
↻ loop until verdict.malicious != null · max 2
→ Prosecutor → verdict
→ Defender → verdict
→ Packager → disposition
→ ask → disposition
Interface
Inputs
reportedEmailobject
The user-reported email: raw source or a message id.
Outputs
dispositionstring
The verdict and, if malicious, the approved remediation package.
Depends on · 4 MCP servers
Multi-account MCP for Gmail, Calendar, Drive, Docs, and Sheets — 61 tools, tokenio.github.adelaidasofia/google-workspace-mcpmailAVerified
Playwrightio.github.microsoft/playwright-mcpbrowserBVerified
Virustotalio.github.BurtTheCoder/virustotalintelAVerified
io.github.slack/slack-mcpnot in the registry yetchat⤓ Export runnable code
# Generated by @socketcat/compiler for target: langgraph
# blueprint: com.socketcat/phishing-investigator v1.0.0 schema: socketcat.dev/blueprint/v0
# This code is yours. Edit it freely. The socketcat_runtime helper is optional and can be vendored.
from typing import TypedDict
from langgraph.graph import StateGraph, START, END
import socketcat_runtime as _rt
class State(TypedDict, total=False):
reportedEmail: object
disposition: object
headerFindings: object
linkFindings: object
fileFindings: object
__loop_5: object
verdict: object
def node_Fork_0(state):
return {}
def node_Join_1(state):
return {}
def node_AgentInvoke_2(state):
return _rt.run_agent("headerAnalyst", state, {"reportedEmail":"reportedEmail"}, "headerFindings")
def node_AgentInvoke_3(state):
return _rt.run_agent("linkAnalyst", state, {"reportedEmail":"reportedEmail"}, "linkFindings")
def node_AgentInvoke_4(state):
return _rt.run_agent("attachmentAnalyst", state, {"reportedEmail":"reportedEmail"}, "fileFindings")
def node_LoopHeader_5(state):
return {}
def node_LoopTick_6(state):
return {"__loop_5": state.get("__loop_5", 0) + 1}
def node_Nop_7(state):
return {}
def node_AgentInvoke_8(state):
return _rt.run_agent("prosecutor", state, {"verdict":"verdict","fileFindings":"fileFindings","linkFindings":"linkFindings","headerFindings":"headerFindings"}, "verdict")
def node_AgentInvoke_9(state):
return _rt.run_agent("defender", state, {"verdict":"verdict","fileFindings":"fileFindings","linkFindings":"linkFindings","headerFindings":"headerFindings"}, "verdict")
def node_AgentInvoke_10(state):
return _rt.run_agent("packager", state, {"verdict":"verdict"}, "disposition")
def node_Ask_11(state):
return _rt.run_ask("Review the phishing verdict and evidence. Approve remediation, or dismiss as benign.", state, "disposition")
def route_LoopHeader_5(state):
count = state.get("__loop_5", 0)
return "exit" if (_rt.cond("verdict.malicious != null", state) or count >= 2) else "loop"
def build():
b = StateGraph(State)
b.add_node("Fork_0", node_Fork_0)
b.add_node("Join_1", node_Join_1)
b.add_node("AgentInvoke_2", node_AgentInvoke_2)
b.add_node("AgentInvoke_3", node_AgentInvoke_3)
b.add_node("AgentInvoke_4", node_AgentInvoke_4)
b.add_node("LoopHeader_5", node_LoopHeader_5)
b.add_node("LoopTick_6", node_LoopTick_6)
b.add_node("Nop_7", node_Nop_7)
b.add_node("AgentInvoke_8", node_AgentInvoke_8)
b.add_node("AgentInvoke_9", node_AgentInvoke_9)
b.add_node("AgentInvoke_10", node_AgentInvoke_10)
b.add_node("Ask_11", node_Ask_11)
b.add_edge(START, "Fork_0")
b.add_edge("Fork_0", "AgentInvoke_2")
b.add_edge("Fork_0", "AgentInvoke_3")
b.add_edge("Fork_0", "AgentInvoke_4")
b.add_edge("Join_1", "AgentInvoke_8")
b.add_edge("AgentInvoke_2", "Join_1")
b.add_edge("AgentInvoke_3", "Join_1")
b.add_edge("AgentInvoke_4", "Join_1")
b.add_conditional_edges("LoopHeader_5", route_LoopHeader_5, {"loop": "AgentInvoke_8", "exit": "Nop_7"})
b.add_edge("LoopTick_6", "LoopHeader_5")
b.add_edge("Nop_7", "AgentInvoke_10")
b.add_edge("AgentInvoke_8", "AgentInvoke_9")
b.add_edge("AgentInvoke_9", "LoopTick_6")
b.add_edge("AgentInvoke_10", "Ask_11")
b.add_edge("Ask_11", END)
return b.compile()
INPUTS = ["reportedEmail"]
OUTPUTS = ["disposition"]
if __name__ == "__main__":
_rt.main(build, INPUTS, OUTPUTS)
▸blueprint.json (the portable format)
{
"id": "com.socketcat/phishing-investigator",
"flow": {
"type": "sequence",
"blocks": [
{
"type": "parallel",
"blocks": [
{
"in": {
"reportedEmail": "reportedEmail"
},
"out": "headerFindings",
"use": "headerAnalyst",
"type": "agent"
},
{
"in": {
"reportedEmail": "reportedEmail"
},
"out": "linkFindings",
"use": "linkAnalyst",
"type": "agent"
},
{
"in": {
"reportedEmail": "reportedEmail"
},
"out": "fileFindings",
"use": "attachmentAnalyst",
"type": "agent"
}
]
},
{
"max": 2,
"type": "loop",
"until": "verdict.malicious != null",
"blocks": [
{
"in": {
"verdict": "verdict",
"fileFindings": "fileFindings",
"linkFindings": "linkFindings",
"headerFindings": "headerFindings"
},
"out": "verdict",
"use": "prosecutor",
"type": "agent"
},
{
"in": {
"verdict": "verdict",
"fileFindings": "fileFindings",
"linkFindings": "linkFindings",
"headerFindings": "headerFindings"
},
"out": "verdict",
"use": "defender",
"type": "agent"
}
]
},
{
"in": {
"verdict": "verdict"
},
"out": "disposition",
"use": "packager",
"type": "agent"
},
{
"out": "disposition",
"type": "ask",
"prompt": "Review the phishing verdict and evidence. Approve remediation, or dismiss as benign."
}
]
},
"tags": [
"security",
"phishing",
"debate",
"human-in-the-loop"
],
"title": "Phishing Investigator",
"agents": {
"defender": {
"model": {
"hint": "reasoning"
},
"title": "Defender",
"output": {
"type": "object",
"required": [
"malicious"
],
"properties": {
"malicious": {
"type": "boolean"
},
"rationale": {
"type": "string"
},
"confidence": {
"type": "string"
}
}
},
"instructions": "Argue that the email is benign, challenging weak or circumstantial evidence. Return an object with a 'malicious' boolean and a confidence level representing where the debate settled. Default to malicious only when the evidence is clear; flag genuine uncertainty for a human."
},
"packager": {
"model": {
"hint": "fast"
},
"title": "Packager",
"tools": [
"chat.post_message"
],
"instructions": "Package the verdict and the evidence into a short case for the security channel. If malicious, list the recommended remediation (quarantine, block sender, reset if credentials were entered). Do not execute any remediation."
},
"prosecutor": {
"model": {
"hint": "reasoning"
},
"title": "Prosecutor",
"instructions": "Argue that the email is malicious, citing the strongest evidence from the three analysts. If you received a defender rebuttal, sharpen or concede specific points honestly."
},
"linkAnalyst": {
"model": {
"hint": "reasoning"
},
"title": "Link Analyst",
"tools": [
"browser.open",
"intel.url_report"
],
"instructions": "Open each link in the sandboxed browser and check its reputation. Report the final landing domain, any credential-harvest form, redirects, and the reputation verdict. Never enter credentials or submit forms."
},
"headerAnalyst": {
"model": {
"hint": "fast"
},
"title": "Header Analyst",
"tools": [
"mail.get_message"
],
"instructions": "Read the email headers. Report SPF, DKIM, and DMARC results, the true sending path, and any display-name or reply-to mismatch. Facts only."
},
"attachmentAnalyst": {
"model": {
"hint": "fast"
},
"title": "Attachment Analyst",
"tools": [
"intel.file_report"
],
"instructions": "Scan each attachment against threat intelligence. Report file type, hash reputation, and any macro or executable content. Do not execute attachments."
}
},
"estate": {
"verdict": {
"type": "object",
"description": "The debate's verdict, including a 'malicious' flag and confidence."
},
"fileFindings": {
"type": "object",
"description": "The attachment analyst's payload scan."
},
"linkFindings": {
"type": "object",
"description": "The link analyst's detonation and reputation results."
},
"headerFindings": {
"type": "object",
"description": "The header analyst's auth and routing read."
}
},
"$schema": "socketcat.dev/blueprint/v0",
"authors": [
{
"url": "https://socketcat.com",
"name": "SocketCat"
}
],
"license": "MIT",
"servers": [
{
"ref": "io.github.adelaidasofia/google-workspace-mcp",
"alias": "mail"
},
{
"ref": "io.github.microsoft/playwright-mcp",
"alias": "browser"
},
{
"ref": "io.github.BurtTheCoder/virustotal",
"alias": "intel"
},
{
"ref": "io.github.slack/slack-mcp",
"alias": "chat"
}
],
"summary": "Analyzes a reported email from three angles in parallel, debates benign vs. malicious to a verdict, and pauses for human remediation.",
"targets": [
"*"
],
"version": "1.0.0",
"interface": {
"inputs": {
"reportedEmail": {
"type": "object",
"description": "The user-reported email: raw source or a message id."
}
},
"outputs": {
"disposition": {
"type": "string",
"description": "The verdict and, if malicious, the approved remediation package."
}
}
},
"extensions": {
"com.autogen": {
"chat": "group"
}
},
"description": "Three analysts run in parallel on a reported email: a header analyst reads the auth and routing, a link analyst opens the URLs in a sandboxed browser and checks their reputation, and an attachment analyst scans payloads against threat intelligence. A prosecutor and a defender then debate the evidence to a verdict, which suppresses false alarms. A confirmed-malicious verdict is packaged with its evidence and paused for a human to approve remediation. Nothing is quarantined without sign-off."
}