GRIM

Security audit MCP server — finds the gaps, exposures, and active compromise that other scanners miss.

Works on any stack. For AI-built apps and existing complex codebases. Point it at a folder, a repo, or an authorized live URL, and get a prioritized, evidence-backed report with fixes.

MCP registry: io.github.AbduljabbarBXR/grim-mcp (mcp-name: io.github.AbduljabbarBXR/grim-mcp)

Status: v0.5.0. v1 validated against a real compromise; v2 adds the planner, SBOM, MITRE ATT&CK tagging, an IoC hash feed, a persistent findings ledger, nested-archive scanning, a delta cache, and parallel scanning. Zero runtime dependencies (Python stdlib only), runs on Linux/macOS/Windows and Termux. 224 tests passing across Python 3.10–3.13.

---

Quickstart

# from the repo root (no install needed)
PYTHONPATH=src python3 -m grim list
PYTHONPATH=src python3 -m grim scan /path/to/app            # one-shot audit (md report)
PYTHONPATH=src python3 -m grim scan backup.tar.gz --format json --out report.json
PYTHONPATH=src python3 -m grim tool audit_exposure --path /path/to/backup.tar.gz
PYTHONPATH=src python3 -m grim diff old.tar.gz new.tar.gz   # drift / active compromise
PYTHONPATH=src python3 -m grim mcp                          # MCP server on stdio

CI

grim ci /path/to/app --fail-on high --format md          # exit 1 when high+ found
grim ci /path/to/app --format sarif --out grim.sarif     # upload SARIF to code scanning

Drift, endpoints, and fixes

grim watch /path/to/app --save                           # store a known-good baseline
grim watch /path/to/app                                  # diff current state vs baseline
grim endpoints /path/to/app                              # route inventory + auth/input risk
grim fix_plan /path/to/app                               # remediation steps + safe diffs
grim malware /path/to/app                                # heuristics + IoC (+ ClamAV/YARA)
grim check_live https://example.com --scope grim.scope.json          # authorized live checks
grim update-feeds --url https://example.com/grim-feed.json   # sync rules + IoCs

Optional real install: pip install -e . (then grim ... works anywhere).

Develop and test

python3 tests/run_all.py     # runs every tests/test_*.py, no pytest needed
ruff check src tests         # optional lint (pip install ruff)
python -m build              # sdist + wheel

CI runs the full suite on Python 3.10, 3.11, 3.12, and 3.13, plus ruff and a build/install smoke test (.github/workflows/ci.yml).

Install from npm or PyPI

grim-mcp bundles the Python engine and runs it as a stdio MCP server (requires Python 3.10+; the npm package also requires Node 18+):

npx -y grim-mcp            # npm
pip install grim-mcp       # PyPI, then: grim mcp
grim version
grim list

Also listed on the official MCP registry (io.github.AbduljabbarBXR/grim-mcp) and Smithery (abdijabarboxer2009/grim-mcp).

Use as MCP server in opencode

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "grim": {
      "type": "local",
      "command": ["npx", "-y", "grim-mcp"],
      "enabled": true
    }
  }
}

Or point directly at a local checkout:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "grim": {
      "type": "local",
      "command": ["python3", "-m", "grim", "mcp"],
      "enabled": true,
      "environment": { "PYTHONPATH": "/path/to/grim/src" }
    }
  }
}

Then the agent can call audit_exposure, scan_code, diff_artifacts, scan, and friends directly while building or reviewing any app.

---

Table of Contents

1. Why GRIM exists
2. Design principles
3. What GRIM is / is not
4. Architecture
5. The four gap classes
6. MCP tool surface
7. Finding schema
8. Engines and auto-updating feeds
9. Safety and authorization model
10. Roadmap
11. Incident pattern
12. Repo layout
13. Tech stack
14. Non-goals and honest limitations
15. Open questions

---

1. Why GRIM exists

AI coding agents can build and modify whole applications in hours. Security tooling did not adapt to that workflow:

• Scanners exist. Orchestration does not. Semgrep, Trivy, gitleaks, nuclei, ClamAV, YARA, OSV — all open source, all excellent, all separate. Nothing runs the right set per stack, normalizes the output, ranks it, explains it in plain language, and hands the AI agent a fix.
• Signature engines are blind to app-logic flaws. A server malware scanner cleans a webshell but cannot see the vulnerable upload handler that keeps writing new ones.
• File-policy failures are invisible to antivirus. Executable files landing in public upload directories is a policy problem, not a virus signature problem, until it is too late.
• Nobody watches drift. The single highest-signal security check for "is this system compromised right now?" is: what changed since the last known-good state? Almost nobody runs it.

One real incident (see the incident pattern) showed how undetected access, credential harvesting, and executed native binaries can persist for a long time before a manual audit finds them. Every artifact was findable by existing engines. No single tool was looking.

GRIM is the looker.

---

2. Design principles

1. Orchestrator, not reinvention. Wrap mature engines. The value is selection, coverage, normalization, prioritization, and remediation — not another regex engine.
2. Evidence over alarms. Every finding carries: file/URL, timestamp, engine, raw evidence, and reproducible command. No "something is wrong somewhere".
3. Read-only by default. File scanning never modifies. Live checks are passive unless an explicit authorization scope enables active probes.
4. AI-native output. Every tool returns machine-stable JSON plus an optional human narrative. Findings include remediation text an AI agent (or a junior dev) can apply directly.
5. Feed-driven freshness. New malware coverage arrives by syncing upstream feeds (OSV, Semgrep registry, nuclei templates, ClamAV, YARA repos) — not by shipping app updates.
6. Any stack. Next.js, Astro, plain HTML, Laravel/PHP, Node/Express, Python/Django, Go, static sites, container images. Detection first, then stack-specific engines.
7. Low-resource capable. Runs on a laptop, a small VPS, or an Android/Termux device for offline audits of downloaded backups and clones.

---

3. What GRIM is / is not

GRIM is	GRIM is not
An MCP server exposing security tools to AI agents	An antivirus product
A security orchestrator over proven engines	A replacement for Imunify360/ClamAV
Code, dependency, exposure, and drift analysis	A guarantee of 100% coverage
Passive live checks + authorized active checks	An exploitation framework
A findings normalizer with fix guidance	A compliance certification tool

---

4. Architecture

┌──────────────────────────────────────────────────────────────┐
│  AI agent (opencode / Claude / Cursor / any MCP client)      │
└──────────────────────────────┬───────────────────────────────┘
                               │  MCP (stdio / http)
┌──────────────────────────────▼───────────────────────────────┐
│                          GRIM CORE                            │
│  detector → planner → runner → normalizer → ranker → report  │
│  policy (scope, safety)      feeds (auto-update)             │
└───────┬───────────────────────────────────────────────┬──────┘
        │                                               │
┌───────▼──────────┐   ┌──────────────────┐   ┌─────────▼──────┐
│ Static engines   │   │ Exposure/Dir     │   │ Live engines   │
│ semgrep, gitleaks│   │ file-policy,     │   │ nuclei, header │
│ trivy, phpstan,  │   │ backup-diff,     │   │ checks, TLS    │
│ bandit, osv.dev  │   │ watch/baseline   │   │ fingerprint    │
└──────────────────┘   └──────────────────┘   └────────────────┘
        │                       │                      │
┌───────▼───────────────────────▼──────────────────────▼───────┐
│                        Malware engines                        │
│            ClamAV (freshclam) · YARA (community repos)        │
└──────────────────────────────────────────────────────────────┘

Core modules

• detector — identifies stack(s) from manifests and file patterns; selects engine plan.
• planner — decides which tools run given target type (repo/folder/live), depth, and policy.
• runner — executes engines with timeouts, resource caps, parallelism, caching of results.
• normalizer — maps every engine output into the GRIM finding schema (§7).
• ranker — severity × confidence × exploitability scoring; dedupes cross-engine findings.
• report — renders Markdown/JSON/SARIF; attaches remediation guidance.
• feeds — updates rule/signature/template feeds (§8).
• policy — authorization scope enforcement and safety gates (§9).

---

5. The four gap classes

GRIM's coverage model. Every tool belongs to one or more:

1. Code-level gaps — unvalidated inputs, unsafe uploads, missing authorization, injection sinks, dangerous patterns. (classic SAST)
2. Dependency gaps — known CVEs in package ecosystems.
3. Exposure gaps — secrets in code, .env/.git/backups/web-executable files reachable, misconfigured headers/cookies, writable public directories.
4. Active compromise — file drift since baseline, known malware signatures, polyglot webshells, unexpected executables, live indicators (suspicious paths responding 200).

---

6. MCP tool surface

v1 (MVP — ship first)

Tool	Purpose	Inputs	Engines
detect_stack	Identify stack(s) and produce an audit plan	path	built-in detectors
audit_deps	Known CVEs in dependencies	path	OSV.dev API, ecosystem lockfiles
scan_secrets	Leaked keys, tokens, .env in tree	path, config?	gitleaks, trufflehog, heuristics
scan_code	Injection, upload, authz, unsafe pattern findings	path, ruleset?	Semgrep (+registry), PHPStan/Psalm, Bandit, eslint-security
audit_exposure	Web-exposed dangerous files in a tree/backup	path (dir, tar, zip)	built-in file-policy engine
report	Unified prioritized report + fixes	findings, format	ranker + renderer

v2/v3 (shipped)

All previously planned v2 tools are implemented: inventory_endpoints, watch, check_live, and malware_scan. See the shipped-features table below.

v3

• fix_plan — turn findings into patch suggestions / PR-ready diffs (emits applyable diffs; PR creation pending)
• Node agent mode — long-running watchdog for live servers without root (PHP/shell cron companion that reports into GRIM)

v2 shipped (0.2.x)

Tool / feature	Purpose
plan	Ordered, explainable audit plan derived from the detected stack
sbom	CycloneDX 1.5 / SPDX 2.3 bill of materials for resolved dependencies
scan_iocs	Match file hashes against a known-bad IoC store (+ EICAR)
update_feeds	Sync the IoC store from a remote JSON feed
ledger	Persistent findings ledger: new / known / reopened / resolved across audits
MITRE ATT&CK	Every finding auto-tagged with technique IDs (e.g. T1505.003)
Nested archives	deep=true streams the archive and descends into nested zip/tar without extracting ordinary files; only inner archives are spilled, bounded, and reported
Delta cache	SHA-256 keyed per-file result cache; unchanged files are not re-scanned
Parallel scanning	Thread-pool SAST across files (workers)
SARIF 2.1.0	--format sarif and the report tool for code-scanning integrations
ci_scan / grim ci	CI gate: exit code by severity threshold (--fail-on)
watch	Persistent baseline + drift detection (save / diff / status)
fix_plan	Ordered remediation steps, with safe unified diffs (git apply-able)
inventory_endpoints	Route inventory (Express, Laravel, Django, Go, Next.js, Astro) with auth/input/risk
Rule feeds	update_feeds syncs SAST rule overlays + IoC hashes from a JSON feed
check_live	Opt-in, scope-gated live checks: headers, cookies, TLS, allowlisted exposed paths
malware_scan	Built-in heuristics + IoC, plus ClamAV/YARA when installed

Limits and truncation

Every cap is overridable via environment variables (0 = unlimited). When a limit is hit, GRIM sets truncated: true, lists the reasons in meta, emits an info finding ("Scan was truncated"), and prints a warning in Markdown reports — so partial results are never silent.

Env var	Default	Covers
GRIM_MAX_ARCHIVE_DEPTH	5	nested-archive recursion depth
GRIM_MAX_ARCHIVE_BYTES	512 MB	total bytes spilled from nested archives
GRIM_MAX_ARCHIVE_ENTRY_BYTES	512 MB	per nested-archive size cap
GRIM_MAX_ENTRIES	600000	entries examined / manifest entries
GRIM_MAX_CONTENT_READS	60000	per-file content reads
GRIM_MAX_FINDINGS	3000	exposure findings
GRIM_MAX_FILES	20000	source files scanned (code/flow)
GRIM_MAX_RULE_MATCHES	10	hits per rule per file
GRIM_MAX_FILE_FINDINGS	200	findings per file
GRIM_MAX_FLOW_FINDINGS	400	flow-analysis findings
GRIM_MAX_CODE_FILE_BYTES	1 MB	per-file code scan size
GRIM_MAX_SECRET_FILE_BYTES	10 MB	per-file secrets scan
GRIM_MAX_SECRET_FINDINGS	800	secrets findings
GRIM_MAX_SECRET_FILES	200000	files secrets-scanned
GRIM_MAX_PACKAGES	3000	dependency packages queried
GRIM_MAX_SECONDS	0 (off)	wall-clock budget per scan
GRIM_SECRET_WORKERS	8	secrets-scan threads
GRIM_OSV_WORKERS	8	OSV request threads
GRIM_OSV_BUDGET_SECONDS	60	total OSV network budget
GRIM_FEEDS_URL	(unset)	rule + IoC feed URL for update_feeds
GRIM_RULES_CACHE	~/.cache/grim/rules.json	synced rule overlay path
GRIM_CLAMAV	(PATH)	ClamAV binary for malware_scan
GRIM_YARA	(PATH)	YARA binary for malware_scan
GRIM_YARA_RULES	(unset)	YARA rules file/dir for malware_scan

The delta cache (~/.cache/grim/code/findings.json) is content- and path-keyed, invalidated by a rules hash, and written atomically; identical files in different paths never share findings. Set use_cache=false on scan_code to bypass it.

Example call

{
  "tool": "audit_exposure",
  "arguments": {
    "path": "/audits/site-backup.tar.gz",
    "checks": ["web-executable", "dotfiles", "exposed-config", "backup-files", "elf-in-public"]
  }
}

{
  "findings": [
    {
      "id": "GRIM-EXPOS-0007",
      "severity": "critical",
      "category": "CWE-434",
      "owasp": "A04:2021",
      "title": "PHP file present in public upload directory",
      "location": { "file": "app/public/uploads/example.php" },
      "evidence": "PHP code in a web-served upload folder",
      "remediation": "Remove file; block PHP execution in upload dirs; fix upload validation server-side",
      "confidence": 0.99,
      "engine": "grim-exposure"
    }
  ]
}

---

7. Finding schema

All engines normalize to this object:

{
  "id": "GRIM-<CLASS>-<NNNN>",
  "severity": "critical | high | medium | low | info",
  "confidence": 0.0,
  "category": "CWE-xxx",
  "owasp": "A01:2021 | ...",
  "title": "short human title",
  "description": "what it is and why it matters",
  "location": {
    "file": "relative/path",
    "line": 0,
    "url": "https://... (live findings)",
    "artifact": "backup.tar.gz (when extracted)"
  },
  "evidence": "raw snippet / header / hash / timestamp — minimal and safe",
  "remediation": "actionable fix, code-level when possible",
  "references": ["https://..."],
  "engine": "name@version",
  "first_seen": "ISO-8601",
  "tags": ["upload", "rce", "active-compromise"]
}

Design rules:

• No false certainty: confidence always present; ranker sorts by severity × confidence.
• Dedupe: same location + category from multiple engines merges into one finding.
• Safe evidence: truncate secrets; never include full key material in reports.

---

8. Engines and auto-updating feeds

Domain	Engine	Feed / update path	Freshness
Dependencies (all ecosystems)	OSV.dev API	live API	real-time
SAST (multi-language)	Semgrep OSS	registry rules (--config auto + pinned sets)	continuous
Secrets	gitleaks / trufflehog	built-in rules + custom GRIM patterns	per release
PHP	PHPStan, Psalm, Enlightn	composer install	per release
JS/TS	eslint-plugin-security, npm audit	npm	continuous
Python	Bandit, pip-audit	PyPI advisories	continuous
Filesystem/containers	Trivy	built-in DB (auto-download)	daily
Live checks	nuclei	templates repo	daily
Malware signatures	ClamAV	freshclam	hours
Malware heuristics	YARA (Neo23x0/signature-base, Elastic, etc.)	git pull	days
File policy + drift	GRIM built-in	GRIM rules file (remote-syncable)	versioned

The auto-update answer: new malware is caught by feeds, not by GRIM releases. GRIM ships the pipeline; ClamAV/YARA/nuclei/Semgrep/OSV ship the ever-fresh detection data. GRIM's own heuristic rules (upload-dir policy, dangerous patterns) are a single versioned rules file that can be hosted remotely and pulled by every installation.

---

9. Safety and authorization model

• Default mode is read-only and local. File scans never write; no network unless a live tool is invoked.
• Live checks are opt-in and require a scope (grim.scope.json, or an inline scope):

{
  "authorization": { "declared_by": "owner or authorized party", "reference": "contract/ticket id" },
  "targets": [
    { "host": "example.com", "mode": "passive",
      "max_requests_per_minute": 30, "paths_allowlist": ["/", "/api/health"] }
  ],
  "deny": ["*/wp-admin/*"]
}

• Passive is the default. check_live reads headers, cookies, and TLS. Active probes run only when the scope target says mode: active and the exact path is in paths_allowlist; denied patterns are never touched.
• No exploitation payloads, ever. Active mode = safe exposure probes, not attacks.
• Rate-limited and allowlisted. Requests are paced by max_requests_per_minute.
• Backups treated as evidence: extraction is isolated and never modifies source archives.

---

10. Roadmap

Phase 0 — spec and fixtures (this document)

• Two sanitized real-world corpora: a compromised app export and a clean baseline
• Golden output files for regression tests

Phase 1 — v1 tools (MVP)

• detect_stack, audit_deps, scan_secrets, scan_code, audit_exposure, report
• CLI mode + MCP server mode
• Finding schema + Markdown/JSON renderers
• Termux-friendly (no root dependencies for v1 tools)

Phase 2 — v2 tools

• Shipped in 0.2.0: plan, sbom (CycloneDX/SPDX), scan_iocs + update_feeds, ledger, MITRE ATT&CK tagging, nested-archive scanning, delta cache, parallel scanning
• Shipped in 0.3.0: SARIF 2.1.0 export, ci_scan / grim ci exit-code gate, production hardening (correct delta cache, unreadable-archive reporting, single-pass code+flow, parallel secrets, deterministic selection, streamed nested spill, wall-clock budget)
• Multi-language SAST + lockfile coverage: Go, Rust, Java, Kotlin, C#, Ruby, Dart
• Shipped in 0.4.0: watch (baselines + drift), fix_plan (remediation + safe diffs), inventory_endpoints, remote rule feeds, bounded per-occurrence rules, cross-platform CI
• Shipped in 0.5.0: check_live (opt-in, scope-gated live checks) and malware_scan (built-in heuristics + IoC, optional ClamAV/YARA adapters)
• Still planned: PR creation from fix_plan, optional node agent for drift alerts

Phase 3 — platform

• Shipped: stdlib test runner, CI matrix (3.10–3.13) with lint + build/install smoke, PyPI-ready metadata (classifiers, urls, LICENSE, MANIFEST)
• Still planned: PR creation from fix_plan (currently emits applyable diffs), remote rules sync shipped via update_feeds
• Optional node agent (PHP/shell, no root) for shared hosting drift alerts
• Dashboard/report hosting (optional paid tier)

Success metric: for the case-study corpus, GRIM v2 must detect 100% of the artifacts found manually, with zero critical false positives on the clean baseline.

---

11. Incident pattern

A pattern seen in real audits that motivated this project (victim, stack, host, and timeline details withheld):

• A web app stored uploads in a public directory and validated file types using an attacker-controllable allowlist — an arbitrary-file-upload leading to remote code execution.
• An attacker used it to write web-executable files into the public upload folder.
• Over time: credential-harvesting scripts read environment files, a downloader installed a remote C2 client, and native executables were downloaded and executed.
• Host malware scanners cleaned some files and left others publicly reachable.
• A manual audit found everything in under an hour by diffing two backup file listings, inspecting suspicious files, and reading logs.

GRIM coverage mapping:

Artifact	Tool that catches it
Attacker-controllable upload type validation	scan_code
Web-executable shells in public upload dir	audit_exposure, malware_scan
Shell stager, ELF binaries in web dir	audit_exposure (web-executable, elf-in-public)
Newly added executables vs old backup	watch / backup diff
Credential-harvesting script pattern	malware_scan (YARA), scan_code
Publicly reachable shells	check_live (authorized exposure probes)
Exposed env/secret files risk	scan_secrets, audit_exposure

---

12. Repo layout

grim/
├── README.md
├── LICENSE
├── pyproject.toml
├── .github/workflows/ci.yml   # test matrix + lint + build
├── docs/
│   └── validation.md          # real-world case study
├── src/grim/
│   ├── __main__.py            # CLI (scan, tool, diff, plan, sbom, ledger, iocs, mcp)
│   ├── tools.py               # MCP tool registry
│   ├── sbom.py                # CycloneDX 1.5 / SPDX 2.3
│   ├── core/                  # detector, planner, findings, ledger, attack, report
│   ├── engines/               # exposure, secrets, codepatterns, flow, deps, diffscan
│   ├── feeds/                 # IoC store + remote feed sync
│   └── mcp/                   # dependency-free stdio MCP server
├── npm/                       # grim-mcp npm launcher (bundles the Python engine)
└── tests/
    ├── run_all.py             # stdlib test runner (CI entry point)
    ├── test_tools.py          # v1 regression
    ├── test_p0_v2.py          # drift, flow
    ├── test_p1_v2.py          # multi-language SAST, lockfiles, secrets
    ├── test_p2_v2.py          # ledger, sbom, ioc, planner, cache, nested archives
    └── test_p3_cli.py         # CLI, MCP, security edge cases

---

## 13. Tech stack

- **Language:** Python 3.11+ (MCP SDK maturity, scanner ecosystem, Termux support)
- **MCP:** official `mcp` Python SDK; stdio transport first, HTTP later
- **Packaging:** `uv`/`pipx` installable; single `grim` entry point; `grim mcp` server mode
- **Engines:** invoked as subprocesses with structured output (`--json` where available);
  adapters isolate version quirks
- **Rules:** YAML for GRIM heuristics; Semgrep YAML for custom code rules
- **Storage:** local cache dir for engine results + baselines (JSON/SQLite)

---

## 14. Non-goals and honest limitations

- **Not 100% coverage.** The promise is: *the four gap classes, with fast drift detection* —
  not "finds every possible exploit".
- **No proof of exploitation.** Detecting a live exploit in progress without server access is
  probabilistic; GRIM reports indicators with confidence levels.
- **No noisy scanners by default.** Aggressive/active scanning is opt-in and authorized.
- **Won't replace host-level security.** Imunify/ClamAV remain; GRIM adds the code- and
  policy-level layers they lack.
- **Shared-hosting reality:** some engines need a CLI environment (Termux/VPS/laptop); the v1
  toolset is designed to run against downloaded copies so live hosting is never required.

---

## 15. Open questions

- Rule hosting: GitHub raw vs. dedicated CDN for `rules/` sync?
- Baseline storage: per-project local vs. optional encrypted remote for `watch` across machines?
- Licensing model: open-core (v1 free, v2+ paid) vs. service-first (audits) while building?
- First vertical to prove out: freelance client audits, AI-IDE integrations, or hosting partners?
- Node agent language for shared hosting: PHP-only (most compatible) or shell + PHP fallback?

---

*GRIM — because the things that get you are the things nobody was looking at.*