pocmap 2.6.2

A modern, AI-friendly CVE PoC discovery tool with structured data models

PocMap

AI-agent-optimized CVE exploit discovery toolkit for bug bounty hunters and security professionals. Find Proof-of-Concept exploits, CTF labs, bug bounty reports, and vulnerability intelligence from a single interface.

Key Features

• Multi-Source Discovery: Queries GitHub, Exploit-DB, Metasploit, Nuclei, CTF labs, and bug bounty platforms simultaneously
• Structured Pydantic Models: All data validated and serialized with full type safety and JSON Schema support
• MCP Server Integration: 22 AI-native tools via Model Context Protocol for Claude Desktop, Cursor, and other AI agents
• Bug Bounty Toolkit: Complete hunter toolkit with checklists, workflows, report templates, prioritization engine, and scope management
• Rich CLI: 13 commands with colorized tables, progress bars, and bulk processing
• Composable Output: table, json, csv, md, and sarif output on read commands, plus a stable exit-code contract for scripting and CI
• CI Security Gate: bulk --fail-on kev|critical|high|epss>=N fails the build on policy matches and emits SARIF 2.1.0 for GitHub code scanning
• Caching & Offline Mode: persistent, TTL'd HTTP response cache and a first-class --offline mode for air-gapped or repeatable runs
• Snapshot Diffs: latest/discover --diff show only what changed since the last identical run
• Concurrent Processing: Thread pool-based bulk CVE processing with exponential backoff retry logic
• Report Generation: Self-contained HTML reports (styled cards and tables, inline CSS, no external assets or JS) plus JSON export
• Security-Hardened: SSRF protection, sandboxed Jinja2 templates, path traversal prevention, input validation

Installation

# From PyPI (CLI + library)
pip install pocmap

# With the MCP server (MCP SDK 2.x — required for pocmap-mcp)
pip install "pocmap[server]"

# From source (editable)
git clone https://github.com/zebbern/pocmap.git
cd pocmap
pip install -e ".[server,dev]"

# Verify
pocmap --version
pocmap-mcp --help    # only after installing with the [server] extra

Every CLI command is also available as python -m pocmap ... if the pocmap script is not on your PATH. The MCP server is also available as python -m pocmap.mcp_server.

Requirements:

• Python 3.10+ (developed/verified on 3.12)
• Core dependencies: pydantic>=2.0, requests>=2.28, urllib3, typer>=0.9, click, rich>=13.0, beautifulsoup4, markdown, jinja2, python-dotenv (see pyproject.toml for the full list and version pins)

Optional:

• GITHUB_API_TOKEN - GitHub PAT for higher rate limits (recommended)
• NVD_API_KEY - NVD API key for increased rate limits
• The [server] extra (MCP SDK mcp>=2.0,<3) is required only for the MCP server / pocmap-mcp

Quick Start

# Look up a single CVE
pocmap lookup CVE-2021-44228

# Show full description and references
pocmap lookup CVE-2021-44228 --description

# Filter PoCs by programming language
pocmap lookup CVE-2021-44228 --language Python

# Process multiple CVEs from a file
pocmap bulk cves.txt --output ./reports

# Search CTF labs for hands-on practice
pocmap labs CVE-2021-44228

# Search bug bounty reports
pocmap bugbounty CVE-2021-44228

# Get CPEs (affected products) for a CVE
pocmap cpes CVE-2021-44228

# Convert CPE to CVEs
pocmap cpe2cve "cpe:2.3:a:apache:log4j:2.0"

# Export JSON schemas for AI agent integration
pocmap schemas --output ./schemas

# Find recently published CVEs from the last 24 hours
pocmap latest

# Find recent critical CVEs with PoCs from the last 7 days
pocmap latest --since 7d --severity critical --only-with-poc

# Discover CVEs affecting a product by name
pocmap discover "Apache Struts"

# Discover CVEs for a specific product version
pocmap discover "Log4j" --version 2.x

# Machine-readable output (any read command): table (default), json, csv, md, sarif
pocmap lookup CVE-2021-44228 --format json
pocmap latest --since 7d --format sarif --output out/

# Use pocmap as a CI gate (exit 6 if any CVE is in CISA KEV)
pocmap bulk cves.txt --format sarif --fail-on kev

# Only show what changed since the last identical run
pocmap latest --since 24h --diff

# Run self-diagnostics and inspect/clear the response cache
pocmap doctor
pocmap cache info
pocmap cache clear

# Serve everything from the local cache (no network)
pocmap --offline lookup CVE-2021-44228

# Show help with all options
pocmap --help

CLI Commands (13)

Command	Purpose
lookup	Look up a single CVE plus discovered PoCs, DB exploits, and labs
bulk	Process many CVEs from a file or stdin; JSON/HTML reports and CI gate
labs	Find CTF labs and vulnerable environments for a CVE
bugbounty	Find bug bounty reports / write-ups for a CVE
cpes	List affected CPE identifiers for a CVE
cpe2cve	List CVE IDs affecting a CPE identifier
readme	Print a GitHub repo's README
schemas	Export JSON schemas for all data models
latest	Find recently published CVEs with exploit intelligence
discover	Discover CVEs affecting a product by name and version
package	Find vulnerabilities in a dependency and the releases that fix them
doctor	Run self-diagnostics (Python, extras, tokens, cache, connectivity)
cache	Inspect (info) and clear (clear) the persistent HTTP cache

Global options (on pocmap itself, before the command): --format/-f {table,json,csv,md,sarif}, --offline, --quiet/-q, --version/-v. Read commands also accept --format/--quiet locally, which override the global value.

Python API

CVE Information Lookup

from pocmap.services.cve_service import CVEService

cve_svc = CVEService()
info = cve_svc.get_cve_info("CVE-2021-44228")

print(info.id)                    # "CVE-2021-44228"
print(info.description)           # Full vulnerability description
print(info.cvss.base_score)       # 10.0
print(info.cvss.severity.value)   # "CRITICAL"
print(info.cvss.vector_string)    # "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H"
print(info.epss)                  # 97.53 (exploitation probability)
print(info.kev_status)            # True (in CISA KEV catalog)
print(info.cwes)                  # ["CWE-77", "CWE-94"]
print(info.vendor)                # "Apache"
print(info.product)               # "Log4j"
print(info.publication_date)      # "2021-12-10"

Exploit Discovery

from pocmap.services.exploit_service import ExploitService

exploit_svc = ExploitService()

# Find all exploits across all sources
exploits = exploit_svc.find_exploits("CVE-2021-44228")
for ex in exploits:
    print(f"[{ex.source.value}] {ex.title}")
    print(f"  URL: {ex.url}")
    print(f"  Language: {ex.language} | Stars: {ex.stars} | Forks: {ex.forks}")

# Filter by programming language
python_pocs = exploit_svc.filter_by_language(exploits, "Python")
go_pocs = exploit_svc.filter_by_language(exploits, "Go")

# Get a GitHub repo's README
readme = exploit_svc.get_readme("https://github.com/example/poc")

Lab Environments

from pocmap.services.lab_service import LabService

lab_svc = LabService()
labs = lab_svc.find_labs("CVE-2021-44228")
for lab in labs:
    print(f"[{lab.platform.value}] {lab.name}: {lab.url}")
    if lab.setup_instructions:
        print(f"  Setup: {lab.setup_instructions}")

Bug Bounty Reports

from pocmap.services.bb_service import BugBountyService

bb_svc = BugBountyService()
reports = bb_svc.find_reports("CVE-2021-44228")
for report in reports:
    print(f"[{report.source.value}] {report.title}")
    print(f"  URL: {report.url} | PoC included: {report.has_poc}")

Report Generation

from pocmap.services.report_service import ReportService

report_svc = ReportService()

# Single CVE report
entry = report_svc.generate_report("CVE-2021-44228")
print(entry.to_json())

# Bulk report with JSON and HTML output
multi = report_svc.generate_bulk_report([
    "CVE-2021-44228",
    "CVE-2023-38408",
    "CVE-2024-21413",
])
report_svc.save_json_report(multi, "./output")
report_svc.save_html_report(multi, "./output")

Schema Export for AI Agents

from pocmap.models import export_schemas

paths = export_schemas("./schemas")
# Generates: CVSSScore.json, CVEInfo.json, Exploit.json,
#            LabEnvironment.json, BugBountyReport.json,
#            CPEInfo.json, RecentExploitResult.json, ReportEntry.json,
#            MultiReport.json, VersionConstraint.json,
#            ProductDiscoveryResult.json,
#            PackageVulnerability.json, PackageDiscoveryResult.json

Bug Bounty Toolkit

The pocmap.bugbounty module provides a comprehensive toolkit for bug bounty hunters:

Structured Checklists

Phase-based checklists with P0-P4 priority levels, completion tracking, and time estimates:

from pocmap.bugbounty import (
    ReconnaissanceChecklist,
    CVEResearchChecklist,
    ExploitationChecklist,
    ReportingChecklist,
)

# Create and track a checklist
checklist = ReconnaissanceChecklist()
checklist.items[0].complete(notes="Subdomain enumeration complete")
print(checklist.completion_status())  # Progress percentage

Methodology Workflows

Structured, repeatable workflows with entry/exit criteria and difficulty ratings:

from pocmap.bugbounty import (
    CVEToBountyWorkflow,        # CVE -> bug bounty pipeline
    ZeroDayHuntingWorkflow,     # Proactive vulnerability discovery
    PatchGapAnalysisWorkflow,   # Patch timing gap exploitation
)

workflow = CVEToBountyWorkflow()
result = workflow.execute_phase("recon", context={"target": "example.com"})

Report Templates

Platform-specific report templates for HackerOne, Bugcrowd, and internal assessments:

from pocmap.bugbounty import HackerOneTemplate, BugcrowdTemplate

template = HackerOneTemplate()
report = template.render(
    cve_data=cve_info,
    impact="Remote code execution achieved via crafted JNDI lookup",
    steps_to_reproduce=[
        "1. Identify Log4j 2.x instance",
        "2. Send crafted payload to vulnerable endpoint",
        "3. Observe DNS callback confirming RCE",
    ],
)

Prioritization Engine

Multi-strategy CVE prioritization with bounty potential estimation:

from pocmap.bugbounty import prioritize_cves, calculate_bounty_potential

# Sort by composite score (CVSS + EPSS + KEV + exploit availability)
sorted_cves = prioritize_cves(cve_list, strategy="composite")

# Or prioritize by specific factors
sorted_cves = prioritize_cves(cve_list, strategy="epss")        # Exploitation probability
sorted_cves = prioritize_cves(cve_list, strategy="kev_first")   # Known exploited first
sorted_cves = prioritize_cves(cve_list, strategy="bounty_potential")

# Estimate bounty potential
for cve in sorted_cves[:10]:
    bounty = calculate_bounty_potential(cve)
    print(f"{cve['id']}: potential=${bounty['estimated_median']}")

Scope Management

Parse and manage bug bounty program scope, match CVEs to in-scope assets:

from pocmap.bugbounty import ScopeManager, Asset

scope = ScopeManager()
scope.add_program(
    platform="hackerone",
    program="example",
    in_scope=["*.example.com", "api.example.com"],
    out_of_scope=["*.internal.example.com"],
)

# Parse scope from file
scope.parse_scope_file("scope.txt")

# Find CVEs affecting in-scope assets
matches = scope.match_cves_to_scope(cve_list)

Playbooks

JSON playbooks for structured workflows:

from pocmap.bugbounty.playbooks import load_playbook, list_playbooks

# List available playbooks
for pb in list_playbooks():
    print(f"{pb['name']}: {pb['description']} ({pb['difficulty']})")

# Load and execute a playbook
playbook = load_playbook("cve-assessment")
for phase in playbook["phases"]:
    print(f"Phase {phase['phase_id']}: {phase['name']}")
    for step in phase["steps"]:
        print(f"  [{step['priority']}] {step['description']}")

Available playbooks:

• cve-assessment: Full CVE assessment workflow with risk scoring and remediation
• rapid-response: Emergency response for critical/KEV CVEs with time-bounded actions
• bb-submission: Complete bug bounty submission pipeline from finding to report

Recent CVE Discovery

Monitor newly published vulnerabilities and filter by severity, exploitability, and time window. Ideal for security briefings, threat intelligence feeds, and proactive vulnerability management.

pocmap latest

# Recent CVEs from the last 24 hours (default)
pocmap latest

# Specify a relative time window
pocmap latest --since 7d
pocmap latest --since 30d
pocmap latest --since 1h

# Explicit date range
pocmap latest --from 2024-01-01 --to 2024-01-31

# Only CVEs with known PoCs on GitHub
pocmap latest --only-with-poc

# Only CISA KEV entries
pocmap latest --kev-only

# Minimum EPSS score filter
pocmap latest --min-epss 50.0

# Filter by severity levels (comma-separated)
pocmap latest --severity critical,high

# Sort results
pocmap latest --sort cve_date     # newest first (default)
pocmap latest --sort severity     # highest severity first
pocmap latest --sort epss         # highest EPSS first

# Limit results and save to JSON
pocmap latest --since 7d --severity critical --only-with-poc --limit 10 --output ./report.json

Options:

Option	Description
--since	Relative time window: 1h, 24h, 7d, 30d
--from	Start date in YYYY-MM-DD format (overrides --since)
--to	End date in YYYY-MM-DD format
--only-with-poc	Only return CVEs with known PoCs on GitHub
--kev-only	Only return CISA Known Exploited Vulnerabilities
--min-epss	Minimum EPSS score (0-100), e.g., 50.0 for EPSS >= 50%
--severity	Comma-separated severities: critical, high, medium, low
--sort	Sort by: cve_date, severity, or epss
--limit	Maximum results (1-100, default: 50)
--output, -o	Save JSON report to file
--diff, --since-last	Show only what changed since the last identical run (added/removed/changed)
--notify <url>	POST a summary of notable CVEs (critical/high or KEV) to a webhook; with --diff, only the delta is sent
--format, -f	Output format: table (default), json, csv, md, sarif
--quiet, -q	Suppress decorative output

Output includes: CVE ID, description, CVSS severity/score, EPSS, KEV status, vendor, product, publication date, PoC availability, and PoC source counts.

Product Discovery

Find all CVEs affecting a specific product without needing a CVE ID. Product names are resolved through the NVD CPE dictionary to canonical vendor:product identifiers, and CVEs are then fetched by CPE applicability match with the version constraint applied by NVD itself.

pocmap discover

# Discover CVEs for a product by name
pocmap discover "Apache Struts"

# With version wildcard
pocmap discover "Log4j" --version 2.x

# Exact version
pocmap discover "nginx" --version 1.20.1

# With vendor hint
pocmap discover "struts" --vendor apache --version 2.x

# Save results to JSON
pocmap discover "Apache Struts" --version 2.x --output ./struts-cves.json

Options:

Option	Description
--version, -v	Version constraint: 2.x, 2.14.1, >= 2.0
--vendor	Vendor name hint: apache, microsoft, google
--limit	Maximum CVEs to analyze (1-100, default: 50)
--output, -o	Save JSON report to file
--diff, --since-last	Show only what changed since the last identical run (added/removed/changed)
--notify <url>	POST a summary of notable CVEs (critical/high or KEV) to a webhook; with --diff, only the delta is sent
--format, -f	Output format: table (default), json, csv, md, sarif
--quiet, -q	Suppress decorative output

How a product name is resolved

1. Alias fast path. A small curated table maps common shorthands to canonical names (see below). A hit skips the dictionary lookup; a miss costs nothing.

2. NVD CPE dictionary. The product name is resolved to every vendor:product pair NVD files CVEs under, ranked by how many CPE entries back each pair. All pairs are searched and the results unioned, because products change hands: nginx resolves to igor_sysoev:nginx (0 CVEs), nginx:nginx (2) and f5:nginx (41), so taking only the top-ranked pair would find almost nothing. At most 5 pairs are queried (a rate-limit bound, not a precision one — NVD allows 5 requests/30s unauthenticated); any dropped pairs are logged, never silently discarded.

   Editions and product families are included. NVD files enterprise software under a separate product per edition, so a query also admits any product in the same vendor's namespace whose name extends the target: "Confluence" reaches atlassian:confluence_server and atlassian:confluence_data_center, "Jira" reaches jira_service_desk, jira_software_data_center and jira_service_management. A third party's lookalike does not — redhat:kubernetes-client is not Kubernetes, and perforce:gliffy is not Confluence. The bias is deliberate: an extra CVE in the list is visible and dismissible, a missing one is neither.

3. Keyword fallback. If the product cannot be resolved at all, discover falls back to NVD full-text search. This is materially weaker — it matches CVE descriptions, so it is both noisy and incomplete. The result reports which path ran:

search_sources	matched_cpes	Meaning
nvd_cpe_match	the resolved CPE prefixes	Authoritative applicability match
nvd_keyword_search	empty	Unresolvable product; noisy full-text fallback

Rate limits. Discovery costs one dictionary lookup plus one query per resolved pair. Unauthenticated NVD allows 5 requests per 30 seconds, so setting NVD_API_KEY is considerably more valuable than it used to be. Responses are cached (see Caching & Offline Mode), so repeat runs are cheap.

Product Alias System

The discovery command recognizes common product aliases and abbreviations, so you don't need to know the exact canonical product name:

Alias Input	Resolved Product
struts	Apache Struts
log4j, log4j2	Log4j
httpd, apache2	Apache HTTP Server
k8s, kube	Kubernetes
postgres, pgsql	PostgreSQL
es, elastic	Elasticsearch
nodejs, node	Node.js
ghe	GitHub Enterprise
ror	Ruby on Rails
wp	WordPress
ie, msie	Internet Explorer

Aliases are matched on the whole name, ignoring separators — apache struts, apache_struts and Apache-Struts are equivalent — and a known vendor phrase is peeled off first, so "Palo Alto PAN-OS" becomes vendor=palo alto + product=pan-os. Matching is deliberately not substring-based: anything the table does not recognize goes to the CPE dictionary, which covers the full NVD catalogue rather than this short list.

Version Constraint Format

Version constraints support multiple formats for flexible version matching:

Format	Example	Meaning
Wildcard	2.x	Any version in major version 2
Exact	2.14.1	Exactly version 2.14.1
Major.Minor	2.14	Version 2.14.x
Range (>=)	>= 2.0	Version 2.0 and above
Range (<=)	<= 1.20	Version 1.20 and below
Range (>)	> 1.0	Above version 1.0
Range (<)	< 3.0	Below version 3.0
None (omit)	-	Any version

Results are grouped into three confidence tiers. Matching considers every (vendor, product) pair a CVE is filed under, not just one — a CVE typically names the vulnerable component plus every distribution that shipped it, and judging it by a single pair misclassifies the component the CVE is actually about:

• Confirmed: Vendor AND product match AND version constraint is met
• Possibly: Vendor OR product matches but version info is unclear
• Not enough data: CVE has insufficient product/version information

Version matching is an interval-overlap test that honours NVD's out-of-band range attributes (versionStartIncluding / versionEndExcluding), which is where modern CVE records express affected ranges — the literal version field in the CPE string is usually just *.

Dependency Vulnerabilities

pocmap discover answers "is the nginx we run vulnerable". pocmap package answers the other half: is the dependency we ship vulnerable, and what do we upgrade to.

The two are keyed differently and are not interchangeable. NVD files vulnerabilities against a CPE product, which is right for deployed software but carries no package coordinate and no fix version. OSV.dev files them against a package (PyPI/django, Maven/org.apache.logging.log4j:log4j-core) and records the releases that fix each one. OSV needs no API key and is not bound by NVD's 5-requests-per-30-seconds limit, so this path stays fast without NVD_API_KEY.

# What's wrong with the version we actually ship, and what fixes it
pocmap package PyPI django --version 3.2.0
pocmap package npm lodash --version 4.17.20
pocmap package Maven org.apache.logging.log4j:log4j-core --version 2.14.1

# Ecosystem names are case-insensitive here and normalized for you
pocmap package pypi requests --version 2.25.0
pocmap package debian:12 nginx

# Only what you can actually act on, as JSON
pocmap package PyPI django --version 3.2.0 --fixable-only --format json

# SARIF for code scanning, in a CI dependency gate
pocmap package npm lodash --version 4.17.20 --format sarif --output out/

Package Vulnerabilities: Maven/org.apache.logging.log4j:log4j-core@2.14.1
Found 7 | with a fix: 7 | no fix published: 0
Severity  CVSS   EPSS  KEV  Advisory        Fixed in
CRITICAL  10.0  100.0  yes  CVE-2021-44228  2.15.0, 2.3.1, 2.12.2
CRITICAL   9.0  100.0  yes  CVE-2021-45046  2.16.0, 2.12.2
HIGH       8.6  100.0       CVE-2021-45105  2.12.3, 2.17.0, 2.3.1

Options:

Option	Description
--version, -v	Installed version. Strongly recommended — OSV then evaluates its own affected ranges and returns only advisories that genuinely apply.
--fixable-only	Only advisories with a published fix. Applied after ranking and --limit, so it narrows the top-N by risk rather than fetching N fixable ones.
--limit	Maximum advisories (1-1000, default: 100), taken from the top of the risk ranking.
--output, -o	Save the JSON report to a file.
--format, -f	table (default), json, csv, md, sarif.
--quiet, -q	Suppress decorative output.

Supported ecosystems. All 50 OSV ecosystems, including PyPI, npm, Go, Maven, crates.io, RubyGems, Packagist, NuGet, Hex, Pub, CRAN, Hackage, ConanCenter, SwiftURL, vcpkg, and the distributions Debian, Ubuntu, Alpine, Red Hat, Rocky Linux, AlmaLinux, SUSE, Wolfi, Chainguard and Bitnami. OSV itself is case-sensitive (pypi is rejected outright); pocmap normalizes common spellings and aliases first, so pypi, cargo, golang, composer and debian:12 all resolve. A release qualifier is preserved: alpine:v3.19 -> Alpine:v3.19.

Reading the results

Ranking is by exploitation risk, not CVSS. Results are ordered CISA KEV first, then EPSS, then CVSS — the top row is what is actually being exploited, which is often not the highest-scoring entry. EPSS and KEV come from bulk catalogues pocmap already caches, so enriching a 100-advisory result costs two cached downloads, not 100 API calls.

Several Fixed in versions is normal. Maintainers backport a fix to each supported branch, so log4j-core is fixed in 2.3.1, 2.12.2 and 2.15.0. Take the one on your own major version.

An empty result is not proof of safety. OSV returns an empty body both for a package it has never heard of and for one with no known issues, so the two are indistinguishable — check the spelling first. This matters most for Maven, where the name must be the full groupId:artifactId; a bare log4j-core matches nothing and looks clean.

Counts describe different things. total_found is how many advisories were found; returned/truncated describe what --limit left. --fixable-only never reports a package as clean — if it empties the list, the output says how many were filtered out.

no fix published means exactly that — the advisory is real and there is no upgrade to take, so it needs a workaround or a risk decision rather than a version bump.

Severity source. A CVSS 3.x vector is scored locally to a base score. CVSS 4.0 scores via a lookup table that pocmap does not implement, so a 4.0-only advisory shows - for CVSS and falls back to the publisher's own rating rather than a plausible-but-wrong number.

Output Formats & Exit Codes

Read commands emit machine-readable output via --format/-f. --format and --quiet/-q can be set globally (before the command) or per command (the local value wins).

pocmap lookup CVE-2021-44228 --format json      # structured view model to stdout
pocmap latest --since 7d --format csv           # spreadsheet-ready rows
pocmap discover "Log4j" --format md             # Markdown table for tickets/wikis
pocmap latest --since 24h --format sarif        # SARIF 2.1.0 for code scanning
pocmap -f json latest --since 7d                # global form

Format	Value	Notes
Table	table	Default. Rich colorized tables (human-facing).
JSON	json	JSON-serializable view model to stdout, nothing else.
CSV	csv	One row per record (csv.DictReader-friendly).
Markdown	md	A Markdown table for tickets/wikis.
SARIF	sarif	SARIF 2.1.0 log for CI code scanning.

Format support by command:

• lookup, doctor, cache info, cache clear: table, json
• labs, bugbounty, cpes, cpe2cve: table, json, csv, md
• latest, discover: table, json, csv, md, sarif
• bulk: table (writes JSON + HTML files), json, csv, sarif

SARIF results are keyed on CVE IDs, so it is available only on the CVE-list commands (latest, discover, bulk). Requesting --format sarif on any other command exits 4 (invalid input) with a clear message. Severity maps to SARIF levels as critical/high -> error, medium -> warning, low -> note; EPSS, KEV, exploit count, and CVSS ride along in result.properties, and each CVE's NVD page is the rule helpUri.

Exit-Code Contract

Every command returns a stable, documented exit code (see src/pocmap/utils/exit_codes.py) so scripts and CI can react to why a command stopped, not just whether it succeeded:

Code	Name	Meaning
0	OK	Success — the command ran and produced output.
1	ERROR	Generic / unclassified error.
2	NO_RESULTS	Ran fine but found nothing (empty result set).
3	NOT_FOUND	Requested resource does not exist upstream (e.g. unknown CVE).
4	INVALID_INPUT	Caller input was malformed (bad CVE ID, unsafe path, bad --fail-on).
5	UPSTREAM_ERROR	An upstream data source failed (network, rate limit, 5xx, offline cache miss).
6	POLICY_FAIL	A bulk --fail-on policy condition matched (the CI gate tripped).

These values are a public contract: existing codes are never renumbered.

Caching & Offline Mode

PocMap keeps a persistent, TTL'd HTTP response cache on disk (default ./.cache). This turns network-bound calls into sub-second cached ones, dodges GitHub/NVD rate limits, and backs a real offline mode. Non-200 and error responses are never cached.

# Warm the cache with a normal (online) run, then work entirely offline
pocmap lookup CVE-2021-44228
pocmap --offline lookup CVE-2021-44228     # served from cache, zero network I/O

# Inspect / clear the cache
pocmap cache info                          # location, entry count, on-disk size
pocmap cache clear                         # delete every cached entry

In --offline mode (or with POCMAP_OFFLINE=1) HTTP GETs are served only from the cache; a cache miss surfaces a clear offline error and exits 5 (UPSTREAM_ERROR) rather than masquerading as "not found" or "no results". An expired-but-cached entry is served stale offline (an air-gapped run cannot refresh it, so stale data beats an error) — only a genuinely absent entry raises. Online runs are unaffected: they still honour the TTL and refetch expired entries.

Cache / offline configuration:

Variable	Default	Description
POCMAP_CACHE_ENABLED	true	Enable/disable the persistent HTTP cache.
POCMAP_CACHE_DIR	./.cache	Directory for cached responses.
POCMAP_CACHE_TTL	3600	Seconds a cached entry stays fresh.
POCMAP_CACHE_MAX_MB	200	On-disk cache cap (MB) before LRU eviction.
POCMAP_OFFLINE	false	Serve only from cache; a miss errors instead of hitting the network.

Verifying PoCs (opt-in)

The CVE indexes list repositories that mention a CVE, which is not the same as repositories that exploit it — link lists, course notes and personal repos all show up. Star count does not separate them either: a popular repo can be an index, and a genuine one-file PoC often has zero stars.

verify_github_pocs downloads the top PoCs' source and reports what each actually contains:

Verdict	Meaning
confirmed	Names the CVE in code and ships runnable code. The only tier that claims the repo exploits the CVE.
likely	Names the CVE but has no code — a writeup.
unverified	Has code, but never names this CVE. Unproven, not disproven: a PoC may be named for its target instead.
unrelated	No mention, or an index — a link list, notes repo or scan dump.

It also derives the language from file extensions, so it needs zero GitHub API calls.

The index test is how many distinct CVEs the repository cites, not how many files it has: a PoC or a writeup is about one vulnerability, while a link list cites dozens. Across a 55-repository sample, genuine PoCs cited at most 3 distinct CVEs and the indexes cited 10, 22 and 118 — so the boundary sits in a wide empty gap. Citing many CVEs only counts against a repo that also ships essentially no code, so a multi-CVE exploit toolkit is not mistaken for a list.

export POCMAP_ALLOW_FETCH_POC_SOURCE=1     # CLI / Python API

Using the MCP server? export will not work. MCP clients launch the server with a filtered environment — the stdio transport inherits only HOME, LOGNAME, PATH, SHELL, TERM and USER — so no POCMAP_* variable set in your shell reaches the server, and verify_github_pocs will keep returning not_enabled. Put it in your client config's env block instead (this applies to GITHUB_API_TOKEN and NVD_API_KEY too):

{
  "mcpServers": {
    "pocmap": {
      "command": "uvx",
      "args": ["--from", "pocmap[server]", "pocmap-mcp"],
      "env": {
        "POCMAP_ALLOW_FETCH_POC_SOURCE": "1",
        "POCMAP_POC_SOURCE_DIR": "/home/you/.local/share/pocmap/poc-source",
        "GITHUB_API_TOKEN": "ghp_xxxxxxxxxxxx"
      }
    }
  }
}

This is off by default and must be set deliberately. It writes third-party exploit code to disk, which endpoint protection will often quarantine — run it on an isolated VM or a dedicated research host. pocmap never executes, imports, or evaluates fetched content; it only reads bytes.

The fetcher is bounded and hardened: owner/repo names are validated before they reach the URL, transfers go through the same SSRF-guarded HTTP client as everything else (per-hop redirect re-validation included), downloads and extracted sizes are both capped so a decompression bomb cannot fill the disk, and archive members that are absolute, contain .., or are symlinks/devices are dropped rather than extracted.

Variable	Default	Description
POCMAP_ALLOW_FETCH_POC_SOURCE	false	Master switch. Nothing is fetched unless this is set.
POCMAP_POC_SOURCE_DIR	<cache_dir>/poc-source	Extraction root.
POCMAP_POC_SOURCE_MAX_MB	100	Per-repo cap on download and extracted size.
POCMAP_POC_SOURCE_TOTAL_MAX_MB	1000	Total cap; oldest fetched repos are evicted first.

Offline mode has nothing to serve here — tarballs deliberately bypass the HTTP response cache — so --offline reports that plainly instead of failing obscurely.

Where to point POCMAP_POC_SOURCE_DIR

The default sits under the response cache, which is convenient but is not the right place on every machine. Two things decide it: whether an on-access scanner will quarantine the files, and whether the directory is synced anywhere.

Windows. Defender scans NTFS on access and will quarantine exploit source, which both interrupts the fetch and silently corrupts any scoring that reads the files back. Either put the directory inside WSL2 — its ext4 lives in a VHDX that Defender does not real-time scan the way it does NTFS — or add a Defender exclusion for a dedicated path:

# Option A: run pocmap inside WSL2 and keep the source there
#   (from your WSL shell)
export POCMAP_POC_SOURCE_DIR="$HOME/.local/share/pocmap/poc-source"

# Option B: stay on Windows and exclude one dedicated directory (admin PowerShell)
New-Item -ItemType Directory -Force "$env:LOCALAPPDATA\pocmap\poc-source"
Add-MpPreference -ExclusionPath "$env:LOCALAPPDATA\pocmap\poc-source"
$env:POCMAP_POC_SOURCE_DIR = "$env:LOCALAPPDATA\pocmap\poc-source"

Exclude the narrowest path that works — one directory used only for this — never your home directory or the whole repo.

Linux / macOS. No on-access scanner by default, so a per-user cache path is fine:

export POCMAP_POC_SOURCE_DIR="$HOME/.cache/pocmap/poc-source"

VM or dedicated research host. Anywhere. This is the intended environment and needs no special handling.

Do not point it at a synced or shared folder. OneDrive, Dropbox, iCloud Drive and Google Drive will upload the exploit source to cloud storage, where the provider's own scanner may flag the account and the content may be shared further than intended. This is easy to hit by accident on Windows, where Documents is frequently redirected to OneDrive — so a repo cloned there gets a OneDrive-synced .cache/ with it. Prefer %LOCALAPPDATA% (never synced) over anything under Documents.

The same reasoning applies to POCMAP_CACHE_DIR, though it holds only API responses rather than exploit code, so it is far less sensitive.

Diagnostics: doctor & cache

pocmap doctor is the fastest path from "installed" to "working". It checks the Python version, the optional [server] extra, the format of your GITHUB_API_TOKEN / NVD_API_KEY (never printing their values), that the cache directory is writable, and runs a live connectivity probe against NVD and the GitHub API. It prints a PASS/WARN/FAIL table and exits nonzero if any check FAILs.

pocmap doctor                  # full run with a live connectivity probe
pocmap doctor --offline        # skip the network probe (labelled SKIPPED)
pocmap doctor --format json    # machine-readable check results

pocmap cache info|clear reports and clears the response cache (see above).

PocMap in CI

bulk is a composable CI gate. Point it at a CVE list (a file, or - to read stdin), choose a machine format, and use --fail-on to fail the build on a policy match:

# Fail the build (exit 6) if any dependency CVE is in the CISA KEV catalog,
# and write a SARIF log for GitHub code scanning.
pocmap bulk cves.txt --format sarif --output out/ --fail-on kev

# Pipe CVE IDs straight from another tool
grep -oE 'CVE-[0-9]{4}-[0-9]+' sbom.txt | pocmap bulk - --format json --fail-on critical

--fail-on accepts critical, high (HIGH or worse), kev, or epss>=N (e.g. epss>=50 on the 0-100 EPSS scale). A match exits 6 (POLICY_FAIL) — distinct from a generic error — so CI can tell a tripped gate apart from an operational failure; no match exits 0. A malformed --fail-on exits 4. In table mode bulk preserves its historical behaviour (writes a JSON and an HTML report to --output); the machine formats (json/csv/sarif) emit a clean stdout summary and write no files, so the stream stays parseable.

See examples/ci-github-actions.yml for a ready-to-use GitHub Actions job that runs the gate and uploads the SARIF to code scanning, and the examples/ directory for more runnable scripts.

AI Agent Integration

PocMap includes a full MCP (Model Context Protocol) server exposing 22 AI-native tools for integration with Claude Desktop, Cursor, and other MCP-compatible clients.

MCP Server Setup for Claude Desktop

Recommended: uv on PATH, no local clone required. --from pocmap[server] pulls the package with the MCP SDK and runs the pocmap-mcp console script over STDIO.

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Windows: %APPDATA%/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "pocmap": {
      "command": "uvx",
      "args": ["--from", "pocmap[server]", "pocmap-mcp"],
      "env": {
        "GITHUB_API_TOKEN": "ghp_xxxxxxxxxxxx",
        "NVD_API_KEY": "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx"
      }
    }
  }
}

Pin a release with pocmap-mcp@X.Y.Z as the last arg (that PyPI version must include the pocmap-mcp entry point). Optional env vars raise GitHub / NVD rate limits.

Already installed locally (pip install "pocmap[server]" or pip install -e ".[server]"):

{
  "mcpServers": {
    "pocmap": {
      "command": "pocmap-mcp",
      "args": [],
      "env": {
        "GITHUB_API_TOKEN": "ghp_xxxxxxxxxxxx",
        "NVD_API_KEY": "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx"
      }
    }
  }
}

Running the MCP Server

The MCP server requires the MCP SDK ([server] extra). The SDK supports protocol revisions up to 2026-07-28 (the stateless core), and negotiates down for clients that have not upgraded. Measured over STDIO — the default, and what Claude Desktop uses — a client performing the initialize handshake currently negotiates 2025-11-25, since 2026-07-28 is reached through the stateless server/discover path rather than the handshake:

pip install "pocmap[server]"
# or from a clone: pip install -e ".[server]"

# STDIO (default — what Claude Desktop / most MCP clients expect)
pocmap-mcp
python -m pocmap.mcp_server

# Other transports / flags
pocmap-mcp --transport sse
pocmap-mcp --transport http --host 0.0.0.0 --port 9000
pocmap-mcp --debug

Repo-root python mcp_server.py is a thin launcher shim to the same module (handy in a git checkout). See also examples/mcp-config.json.

MCP Tools (22 Total)

Tool	Category	Description
lookup_cve	CVE Intel	Full CVE details from NVD, CVE.org, CISA KEV, EPSS
get_epss_score	CVE Intel	EPSS exploitation probability score (0.0-1.0) with risk level
check_kev_status	CVE Intel	Check CISA Known Exploited Vulnerabilities catalog status
get_attack_techniques	CVE Intel	MITRE ATT&CK techniques a CVE maps to — how it's exploited and what follows
find_github_pocs	Exploits	GitHub PoC repos with stars, language, and forks
verify_github_pocs	Exploits	Reads PoC source to score whether a repo really exploits the CVE (opt-in)
find_metasploit_module	Exploits	Metasploit module availability and msfconsole command
find_exploitdb_entry	Exploits	ExploitDB entry with searchsploit command
find_nuclei_template	Exploits	Nuclei scanner template for detection/verification
find_bug_bounty_reports	Research	Bug bounty write-ups from HackerOne, PentesterLand
find_practice_labs	Labs	CTF labs on Vulhub and HackTheBox
find_vulhub_docker	Labs	Vulhub Docker Compose environment with setup steps
find_recent_exploits	Discovery	Recently published CVEs with PoC/KEV/severity filters
discover_product_cves	Discovery	Find CVEs by product name with version constraints
discover_package_cves	Discovery	Dependency vulnerabilities + the releases that fix them (OSV, no API key)
cve_to_cpe	Conversion	Convert CVE to affected CPE identifiers
cpe_to_cve	Conversion	Find all CVEs affecting a given product (CPE)
generate_json_report	Reports	One-shot CVE assessment — details + all exploits + labs + bug bounty reports for one or many CVEs in a single call
generate_html_report	Reports	Self-contained HTML report with styled cards
get_cve_assessment_playbook	Playbooks	Full CVE assessment workflow playbook
get_rapid_response_playbook	Playbooks	Emergency response playbook for critical CVEs
get_bug_bounty_playbook	Playbooks	Bug bounty submission workflow playbook

MCP Resources

Resource	URI Pattern	Content
CVE Info	cve://{cve_id}	Full CVE details as human-readable text
Exploits	exploits://{cve_id}	All available exploits and PoCs
Report	report://{cve_id}	Generated vulnerability report (JSON)

MCP Prompts

Prompt	Description
vulnerability_assessment	5-phase structured CVE assessment workflow
exploit_research	Deep exploit analysis with detection engineering focus
bug_bounty_analysis	Real-world impact analysis from bug bounty write-ups

Example Agent Workflow

User: "Should I prioritize CVE-2021-44228, CVE-2023-38408, or CVE-2024-21413?"

Agent:
1. lookup_cve("CVE-2021-44228")     -> CVSS 10.0 CRITICAL, EPSS 0.9753, KEV=true
2. lookup_cve("CVE-2023-38408")     -> CVSS 9.8 CRITICAL, EPSS 0.3124, KEV=true
3. lookup_cve("CVE-2024-21413")     -> CVSS 8.8 HIGH, EPSS 0.8912, KEV=true
4. get_epss_score for each          -> Confirm exploitation probabilities
5. find_github_pocs for each        -> Count available exploits
6. check_kev_status for each        -> Confirm KEV status
7. Prioritize: Log4j (highest EPSS + most exploits) > CVE-2024-21413 > CVE-2023-38408

JSON Schemas

Export structured JSON schemas for all data models to integrate with AI agents, validation pipelines, and external tools:

from pocmap.models import export_schemas

paths = export_schemas("./schemas")
for p in paths:
    print(f"  {p.name}")
# CVSSScore.json
# CVEInfo.json
# Exploit.json
# LabEnvironment.json
# BugBountyReport.json
# CPEInfo.json
# RecentExploitResult.json
# ReportEntry.json
# MultiReport.json
# VersionConstraint.json
# ProductDiscoveryResult.json
# PackageVulnerability.json
# PackageDiscoveryResult.json

Use these schemas for:

• AI Agent Context: Provide schema files to AI agents so they understand data structures
• Validation Pipelines: Validate incoming/outgoing data against schemas
• API Documentation: Auto-generate API docs from schemas
• Type Generation: Generate TypeScript, Go, or Rust types from JSON schemas

Architecture

+------------------+     +------------------+     +------------------+
|     CLI Layer    |     |   MCP Server     |     |   Python API     |
|   (Typer/Rich)   |     |  (MCP SDK / 21   |     |   (Services)     |
+------------------+     |     Tools)       |     +------------------+
         |               +------------------+             |
         |                         |                      |
         v                         v                      v
+------------------+     +------------------+     +------------------+
|   Service Layer  |<--->|   Service Layer  |<--->|   Service Layer  |
|                  |     |                  |     |                  |
|  CVEService      |     |  ExploitService  |     |  ReportService   |
|  BugBountyService|     |  LabService      |     |  + 2 more        |
+------------------+     +------------------+     +------------------+
         |                         |                      |
         v                         v                      v
+------------------+     +------------------+     +------------------+
|  Client Layer    |     |  Client Layer    |     |   Models Layer   |
|                  |     |                  |     |                  |
|  NVDClient       |     |  GitHubClient    |     |  CVEInfo         |
|  CPEDictClient   |     |  ExploitClient   |     |  Exploit         |
|  CVEOrgClient    |     |  + others        |     |  + 11 more       |
+------------------+     +------------------+     +------------------+
         |                         |
         v                         v
+-------------------------------------------------------------+
|                    External Data Sources                     |
|  NVD API  CVE.org  CISA KEV  EPSS  GitHub  ExploitDB  etc.  |
+-------------------------------------------------------------+

Layered architecture:

1. Presentation Layer: CLI (cli.py) + MCP Server (pocmap.mcp_server / pocmap-mcp)
2. Service Layer: Business logic (7 services: CVE, Exploit, Lab, Report, Bug Bounty, Recent, Product Discovery)
3. Client Layer: External API clients (NVD, GitHub, CVE.org, ExploitDB, etc.)
4. Model Layer: 16 Pydantic models with full validation and JSON Schema support (13 exported as standalone JSON schemas)
5. Utility Layer: HTTP client with retries, formatters, validators, config
6. Toolkit Layer: Bug bounty hunter toolkit (checklists, methodology, templates, prioritization, scope, automation)

Security Features

SSRF Protection

All HTTP requests pass through is_safe_url() validation that blocks:

• Internal hosts: localhost, 127.0.0.1, 0.0.0.0, ::1 (matched by exact host / dotted-suffix, so public hosts that merely contain those strings — e.g. the IPv6 literal 2606:4700:4700::1111 — are not falsely blocked)
• Cloud metadata endpoints: 169.254.169.254 (AWS), metadata.google.internal (GCP), 100.100.100.200 (Alibaba)
• Private IP ranges, loopback, link-local, and reserved addresses
• Numeric-encoded IPs (decimal/hex/octal) and IPv4-mapped IPv6 that canonicalize to an internal address
• Non-HTTP(S) schemes: file://, ftp://, gopher://, dict://

Redirects are followed manually and every hop is re-validated through the same guard, and credential-bearing headers (Authorization, Cookie, NVD apiKey) are stripped on a cross-origin redirect so a token is never replayed to a redirect target.

CSV Injection Prevention

CSV export neutralizes spreadsheet formula injection (CWE-1236): a string cell that begins with a formula character (=, +, -, @, tab, or CR) is prefixed with a single quote so externally-sourced text (CVE descriptions, repo names) cannot execute when the file is opened in Excel / Google Sheets. Genuine numbers are left intact.

Sandboxed Templates

Jinja2 templates use SandboxedEnvironment with BaseLoader (no filesystem access) and select_autoescape for HTML/XML contexts. Prevents Server-Side Template Injection (SSTI) attacks.

Path Traversal Protection

File operations use safe_path() which normalizes paths and validates they stay within the base directory. Raises ValueError on traversal attempts.

Input Validation

• CVE IDs validated against ^CVE-\d{4}-\d+$ regex pattern
• CPE strings parsed with strict format validation
• Maximum bulk size limit (100 CVEs) prevents DoS
• All inputs sanitized before external API calls

XSS Prevention

• HTML report generation uses auto-escaped template rendering
• All user-facing output is properly escaped
• No inline JavaScript execution in generated HTML reports

Configuration

Configuration is loaded from environment variables (prefixed with POCMAP_) and optional .env file:

# Create .env file
cat > .env << 'EOF'
GITHUB_API_TOKEN=ghp_xxxxxxxxxxxxxxxxxxxx
NVD_API_KEY=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
POCMAP_HTTP_TIMEOUT=30
POCMAP_MAX_RETRIES=3
POCMAP_BACKOFF_FACTOR=1.5
POCMAP_THREAD_POOL_SIZE=10
POCMAP_LOG_LEVEL=INFO
POCMAP_CACHE_ENABLED=true
POCMAP_CACHE_TTL=3600
POCMAP_CACHE_MAX_MB=200
EOF

Variable	Default	Description
GITHUB_API_TOKEN	None	GitHub personal access token for higher rate limits
NVD_API_KEY	None	NVD API key for increased rate limits
POCMAP_HTTP_TIMEOUT	30	HTTP request timeout in seconds
POCMAP_MAX_RETRIES	3	Maximum retry attempts for failed requests
POCMAP_BACKOFF_FACTOR	1.5	Exponential backoff multiplier
POCMAP_THREAD_POOL_SIZE	10	Worker thread count for bulk operations
POCMAP_LOG_LEVEL	INFO	Logging verbosity (DEBUG, INFO, WARNING, ERROR)
POCMAP_CACHE_ENABLED	true	Enable the persistent HTTP response cache
POCMAP_CACHE_DIR	./.cache	Directory for cached responses
POCMAP_CACHE_TTL	3600	Seconds a cached entry stays fresh
POCMAP_CACHE_MAX_MB	200	On-disk cache cap (MB) before LRU eviction
POCMAP_OFFLINE	false	Serve HTTP only from cache; a miss errors instead of hitting the network
POCMAP_ALLOW_FETCH_POC_SOURCE	false	Opt in to downloading PoC source code to disk (see below)
POCMAP_POC_SOURCE_DIR	<cache>/poc-source	Where fetched PoC source is extracted
POCMAP_POC_SOURCE_MAX_MB	100	Per-repo cap, applied to download and extracted size
POCMAP_POC_SOURCE_TOTAL_MAX_MB	1000	Total on-disk cap for fetched sources

See Caching & Offline Mode and the exit-code contract for how these behave at runtime.

Contributing

Adding New Exploit Sources

New exploit sources can be registered via the plugin pattern:

1. Create a new client in src/pocmap/clients/:

# src/pocmap/clients/my_source_client.py
from pocmap.models import Exploit, ExploitSource

class MySourceClient:
    """Client for My Exploit Source."""

    SOURCE = ExploitSource.OTHER  # or add to enum

    def search(self, cve_id: str) -> list[Exploit]:
        # Implement search logic
        return []

2. Integrate into ExploitService in src/pocmap/services/exploit_service.py:

from pocmap.clients.my_source_client import MySourceClient

class ExploitService:
    def __init__(self):
        self._my_source = MySourceClient()

    def find_exploits(self, cve_id: str) -> list[Exploit]:
        exploits = []
        exploits.extend(self._my_source.search(cve_id))
        # ... existing sources
        return exploits

3. Add tests and documentation.

Third-Party Exploit Sources (plugins — no fork needed)

External packages can add exploit sources without modifying pocmap by registering an entry point in the pocmap.exploit_sources group. A source is any object exposing search(cve_id: str) -> list[Exploit] (the ExploitSourcePlugin protocol):

# your package's pyproject.toml
[project.entry-points."pocmap.exploit_sources"]
my-source = "my_pkg.source:MySource"

# my_pkg/source.py
from pocmap.models import Exploit, ExploitSource

class MySource:
    source = "my-source"

    def search(self, cve_id: str) -> list[Exploit]:
        return [Exploit(source=ExploitSource.OTHER, url="https://…", title="…")]

pip install your package and its results automatically appear in pocmap lookup and ExploitService.find_exploits. Plugins are error-isolated: a failing plugin degrades to a FetchStatus.ERROR (visible via find_exploits_with_status) without affecting the built-in sources. A complete runnable example is in examples/example-exploit-source/. Note: entry-point plugins execute third-party code you chose to install — pocmap only calls their search() and aggregates the results with per-source status isolation.

Development Setup

git clone https://github.com/zebbern/pocmap.git
cd pocmap
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Run tests
pytest -v

# Run type checker
mypy src/pocmap

# Run linter
ruff check src/pocmap

License

MIT License - See LICENSE file for details.

---

PocMap is not a weapon. It is a research and defensive tool for security professionals and bug bounty hunters. Always operate within the bounds of applicable law and program scope.