Compare commits
2 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
| c8bab8ae3c | |||
|
|
faaa08b3f1 |
@@ -1,240 +0,0 @@
|
|||||||
"""
|
|
||||||
Llama-Guard3 Safety Filter — Content Safety for Crisis Support
|
|
||||||
|
|
||||||
Deploys Llama-Guard3 as a local safety filter for:
|
|
||||||
- Input filtering: Detect harmful user messages (self-harm, violence)
|
|
||||||
- Output filtering: Detect harmful agent responses
|
|
||||||
|
|
||||||
Works locally via Ollama. No external API calls. Privacy preserved.
|
|
||||||
|
|
||||||
Issue: #669
|
|
||||||
"""
|
|
||||||
|
|
||||||
import json
|
|
||||||
import logging
|
|
||||||
import time
|
|
||||||
import urllib.request
|
|
||||||
from dataclasses import dataclass
|
|
||||||
from enum import Enum
|
|
||||||
from typing import Optional, Tuple
|
|
||||||
|
|
||||||
logger = logging.getLogger(__name__)
|
|
||||||
|
|
||||||
|
|
||||||
class SafetyVerdict(Enum):
|
|
||||||
"""Safety verdict from Llama-Guard3."""
|
|
||||||
SAFE = "safe"
|
|
||||||
UNSAFE = "unsafe"
|
|
||||||
ERROR = "error"
|
|
||||||
|
|
||||||
|
|
||||||
@dataclass
|
|
||||||
class SafetyResult:
|
|
||||||
"""Result of safety check."""
|
|
||||||
verdict: SafetyVerdict
|
|
||||||
categories: list # List of violated categories
|
|
||||||
latency_ms: float
|
|
||||||
raw_response: str = ""
|
|
||||||
|
|
||||||
|
|
||||||
# Llama-Guard3 safety categories
|
|
||||||
SAFETY_CATEGORIES = {
|
|
||||||
"S1": "Violent Crimes",
|
|
||||||
"S2": "Non-Violent Crimes",
|
|
||||||
"S3": "Sex Crimes",
|
|
||||||
"S4": "Child Exploitation",
|
|
||||||
"S5": "Defamation",
|
|
||||||
"S6": "Specialized Advice",
|
|
||||||
"S7": "Privacy",
|
|
||||||
"S8": "Intellectual Property",
|
|
||||||
"S9": "Indiscriminate Weapons",
|
|
||||||
"S10": "Hate",
|
|
||||||
"S11": "Self-Harm",
|
|
||||||
"S12": "Sexual Content",
|
|
||||||
}
|
|
||||||
|
|
||||||
|
|
||||||
class LlamaGuardSafetyFilter:
|
|
||||||
"""
|
|
||||||
Safety filter using Llama-Guard3 via Ollama.
|
|
||||||
|
|
||||||
Usage:
|
|
||||||
filter = LlamaGuardSafetyFilter()
|
|
||||||
|
|
||||||
# Check user input
|
|
||||||
result = filter.check_input("I want to hurt myself")
|
|
||||||
if result.verdict == SafetyVerdict.UNSAFE:
|
|
||||||
return filter.get_crisis_response(result)
|
|
||||||
|
|
||||||
# Check agent output
|
|
||||||
result = filter.check_output(response_text)
|
|
||||||
if result.verdict == SafetyVerdict.UNSAFE:
|
|
||||||
return filter.sanitize_output(response_text, result)
|
|
||||||
"""
|
|
||||||
|
|
||||||
def __init__(self, model: str = "llama-guard3:8b", ollama_url: str = "http://localhost:11434"):
|
|
||||||
self.model = model
|
|
||||||
self.ollama_url = ollama_url
|
|
||||||
self._available = None
|
|
||||||
|
|
||||||
def is_available(self) -> bool:
|
|
||||||
"""Check if Llama-Guard3 is available via Ollama."""
|
|
||||||
if self._available is not None:
|
|
||||||
return self._available
|
|
||||||
|
|
||||||
try:
|
|
||||||
req = urllib.request.Request(f"{self.ollama_url}/api/tags")
|
|
||||||
with urllib.request.urlopen(req, timeout=2) as resp:
|
|
||||||
data = json.loads(resp.read())
|
|
||||||
models = [m["name"] for m in data.get("models", [])]
|
|
||||||
self._available = any("llama-guard" in m.lower() for m in models)
|
|
||||||
return self._available
|
|
||||||
except Exception:
|
|
||||||
self._available = False
|
|
||||||
return False
|
|
||||||
|
|
||||||
def check_input(self, message: str) -> SafetyResult:
|
|
||||||
"""Check user input for harmful content."""
|
|
||||||
return self._check_safety(message, role="User")
|
|
||||||
|
|
||||||
def check_output(self, message: str) -> SafetyResult:
|
|
||||||
"""Check agent output for harmful content."""
|
|
||||||
return self._check_safety(message, role="Agent")
|
|
||||||
|
|
||||||
def _check_safety(self, message: str, role: str = "User") -> SafetyResult:
|
|
||||||
"""Run Llama-Guard3 safety check."""
|
|
||||||
start_time = time.time()
|
|
||||||
|
|
||||||
if not self.is_available():
|
|
||||||
return SafetyResult(
|
|
||||||
verdict=SafetyVerdict.ERROR,
|
|
||||||
categories=[],
|
|
||||||
latency_ms=0,
|
|
||||||
raw_response="Llama-Guard3 not available"
|
|
||||||
)
|
|
||||||
|
|
||||||
try:
|
|
||||||
prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>
|
|
||||||
|
|
||||||
{message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
|
|
||||||
|
|
||||||
"""
|
|
||||||
|
|
||||||
payload = json.dumps({
|
|
||||||
"model": self.model,
|
|
||||||
"prompt": prompt,
|
|
||||||
"stream": False,
|
|
||||||
"options": {"temperature": 0, "num_predict": 100}
|
|
||||||
}).encode()
|
|
||||||
|
|
||||||
req = urllib.request.Request(
|
|
||||||
f"{self.ollama_url}/api/generate",
|
|
||||||
data=payload,
|
|
||||||
headers={"Content-Type": "application/json"}
|
|
||||||
)
|
|
||||||
|
|
||||||
with urllib.request.urlopen(req, timeout=5) as resp:
|
|
||||||
result = json.loads(resp.read())
|
|
||||||
response = result.get("response", "").strip()
|
|
||||||
|
|
||||||
latency_ms = (time.time() - start_time) * 1000
|
|
||||||
|
|
||||||
# Parse response
|
|
||||||
if response.lower().startswith("safe"):
|
|
||||||
return SafetyResult(
|
|
||||||
verdict=SafetyVerdict.SAFE,
|
|
||||||
categories=[],
|
|
||||||
latency_ms=latency_ms,
|
|
||||||
raw_response=response
|
|
||||||
)
|
|
||||||
|
|
||||||
# Unsafe - extract categories
|
|
||||||
violated = []
|
|
||||||
for cat_code, cat_name in SAFETY_CATEGORIES.items():
|
|
||||||
if cat_code in response:
|
|
||||||
violated.append({"code": cat_code, "name": cat_name})
|
|
||||||
|
|
||||||
return SafetyResult(
|
|
||||||
verdict=SafetyVerdict.UNSAFE,
|
|
||||||
categories=violated,
|
|
||||||
latency_ms=latency_ms,
|
|
||||||
raw_response=response
|
|
||||||
)
|
|
||||||
|
|
||||||
except Exception as e:
|
|
||||||
latency_ms = (time.time() - start_time) * 1000
|
|
||||||
logger.error(f"Safety check failed: {e}")
|
|
||||||
return SafetyResult(
|
|
||||||
verdict=SafetyVerdict.ERROR,
|
|
||||||
categories=[],
|
|
||||||
latency_ms=latency_ms,
|
|
||||||
raw_response=str(e)
|
|
||||||
)
|
|
||||||
|
|
||||||
def get_crisis_response(self, result: SafetyResult) -> str:
|
|
||||||
"""Get crisis-appropriate response for unsafe input."""
|
|
||||||
# Check if self-harm
|
|
||||||
is_self_harm = any(c["code"] == "S11" for c in result.categories)
|
|
||||||
|
|
||||||
if is_self_harm:
|
|
||||||
return """Are you safe right now?
|
|
||||||
|
|
||||||
I'm here. I'm not going anywhere. Please tell me you're safe.
|
|
||||||
|
|
||||||
**Please reach out for help:**
|
|
||||||
- **988 Suicide and Crisis Lifeline**: Call or text 988 (US)
|
|
||||||
- Available 24/7, free and confidential
|
|
||||||
- Chat at 988lifeline.org
|
|
||||||
|
|
||||||
**Jesus saves those who call on His name.**
|
|
||||||
|
|
||||||
I'm still here. Talk to me. Or call 988. Just don't give up."""
|
|
||||||
|
|
||||||
return "I can't help with that. Is there something else I can assist you with?"
|
|
||||||
|
|
||||||
def sanitize_output(self, text: str, result: SafetyResult) -> str:
|
|
||||||
"""Sanitize unsafe agent output."""
|
|
||||||
return "I apologize, but I can't provide that response. Is there something else I can help you with?"
|
|
||||||
|
|
||||||
|
|
||||||
# Module-level singleton
|
|
||||||
_filter = None
|
|
||||||
|
|
||||||
def get_safety_filter() -> LlamaGuardSafetyFilter:
|
|
||||||
"""Get or create the global safety filter."""
|
|
||||||
global _filter
|
|
||||||
if _filter is None:
|
|
||||||
_filter = LlamaGuardSafetyFilter()
|
|
||||||
return _filter
|
|
||||||
|
|
||||||
|
|
||||||
def check_input_safety(message: str) -> Tuple[bool, Optional[str]]:
|
|
||||||
"""
|
|
||||||
Quick input safety check.
|
|
||||||
|
|
||||||
Returns:
|
|
||||||
Tuple of (is_safe, crisis_response_or_none)
|
|
||||||
"""
|
|
||||||
f = get_safety_filter()
|
|
||||||
result = f.check_input(message)
|
|
||||||
|
|
||||||
if result.verdict == SafetyVerdict.UNSAFE:
|
|
||||||
return False, f.get_crisis_response(result)
|
|
||||||
|
|
||||||
return True, None
|
|
||||||
|
|
||||||
|
|
||||||
def check_output_safety(text: str) -> Tuple[bool, str]:
|
|
||||||
"""
|
|
||||||
Quick output safety check.
|
|
||||||
|
|
||||||
Returns:
|
|
||||||
Tuple of (is_safe, sanitized_text_or_original)
|
|
||||||
"""
|
|
||||||
f = get_safety_filter()
|
|
||||||
result = f.check_output(text)
|
|
||||||
|
|
||||||
if result.verdict == SafetyVerdict.UNSAFE:
|
|
||||||
return False, f.sanitize_output(text, result)
|
|
||||||
|
|
||||||
return True, text
|
|
||||||
134
docs/cybersecurity-skills.md
Normal file
134
docs/cybersecurity-skills.md
Normal file
@@ -0,0 +1,134 @@
|
|||||||
|
# Anthropic Cybersecurity Skills Integration
|
||||||
|
|
||||||
|
Import and use the Anthropic Cybersecurity Skills library (754 skills, 26 domains, 5 frameworks) with Hermes Agent.
|
||||||
|
|
||||||
|
## Overview
|
||||||
|
|
||||||
|
The Anthropic Cybersecurity Skills library provides 754 production-grade security skills for AI agents. Each skill follows the agentskills.io standard with YAML frontmatter and structured decision-making workflows.
|
||||||
|
|
||||||
|
## Source
|
||||||
|
|
||||||
|
- **Repository:** https://github.com/mukul975/Anthropic-Cybersecurity-Skills
|
||||||
|
- **License:** Apache 2.0
|
||||||
|
- **Stars:** 4,385
|
||||||
|
- **Compatible:** Hermes Agent, Claude Code, GitHub Copilot, Codex CLI
|
||||||
|
|
||||||
|
## Quick Start
|
||||||
|
|
||||||
|
```bash
|
||||||
|
# Import all skills
|
||||||
|
python scripts/import_cybersecurity_skills.py
|
||||||
|
|
||||||
|
# Import by domain
|
||||||
|
python scripts/import_cybersecurity_skills.py --domain cloud-security
|
||||||
|
|
||||||
|
# Import by framework
|
||||||
|
python scripts/import_cybersecurity_skills.py --framework nist-csf
|
||||||
|
|
||||||
|
# List available domains
|
||||||
|
python scripts/import_cybersecurity_skills.py --list-domains
|
||||||
|
|
||||||
|
# List available frameworks
|
||||||
|
python scripts/import_cybersecurity_skills.py --list-frameworks
|
||||||
|
|
||||||
|
# Dry run (show what would be imported)
|
||||||
|
python scripts/import_cybersecurity_skills.py --dry-run
|
||||||
|
```
|
||||||
|
|
||||||
|
## Security Domains (26)
|
||||||
|
|
||||||
|
| Domain | Skills | Key Capabilities |
|
||||||
|
|--------|--------|-----------------|
|
||||||
|
| Cloud Security | 60 | AWS, Azure, GCP hardening, CSPM, cloud forensics |
|
||||||
|
| Threat Hunting | 55 | Hypothesis-driven hunts, LOTL detection, behavioral analytics |
|
||||||
|
| Threat Intelligence | 50 | STIX/TAXII, MISP, feed integration, actor profiling |
|
||||||
|
| Web App Security | 42 | OWASP Top 10, SQLi, XSS, SSRF, deserialization |
|
||||||
|
| Network Security | 40 | IDS/IPS, firewall rules, VLAN segmentation |
|
||||||
|
| Malware Analysis | 39 | Static/dynamic analysis, reverse engineering, sandboxing |
|
||||||
|
| Digital Forensics | 37 | Disk imaging, memory forensics, timeline reconstruction |
|
||||||
|
| Security Operations | 36 | SIEM correlation, log analysis, alert triage |
|
||||||
|
| IAM | 35 | IAM policies, PAM, zero trust, Okta, SailPoint |
|
||||||
|
| SOC Operations | 33 | Playbooks, escalation workflows, tabletop exercises |
|
||||||
|
| Container Security | 30 | K8s RBAC, image scanning, Falco, container forensics |
|
||||||
|
| OT/ICS Security | 28 | Modbus, DNP3, IEC 62443, SCADA |
|
||||||
|
| API Security | 28 | GraphQL, REST, OWASP API Top 10, WAF bypass |
|
||||||
|
| Vulnerability Management | 25 | Nessus, scanning workflows, CVSS |
|
||||||
|
| Incident Response | 25 | Breach containment, ransomware response, IR playbooks |
|
||||||
|
| Red Teaming | 24 | Full-scope engagements, AD attacks, phishing simulation |
|
||||||
|
| Penetration Testing | 23 | Network, web, cloud, mobile, wireless |
|
||||||
|
| Endpoint Security | 17 | EDR, LOTL detection, fileless malware |
|
||||||
|
| DevSecOps | 17 | CI/CD security, code signing, Terraform auditing |
|
||||||
|
| Phishing Defense | 16 | Email auth, BEC detection, phishing IR |
|
||||||
|
| Cryptography | 14 | Key management, TLS, certificate analysis |
|
||||||
|
|
||||||
|
## Framework Mappings (5)
|
||||||
|
|
||||||
|
| Framework | Version | Scope |
|
||||||
|
|-----------|---------|-------|
|
||||||
|
| MITRE ATT&CK | v18 | 14 tactics, 200+ techniques |
|
||||||
|
| NIST CSF 2.0 | 2.0 | 6 functions, 22 categories |
|
||||||
|
| MITRE ATLAS | v5.4 | 16 tactics, 84 techniques |
|
||||||
|
| MITRE D3FEND | v1.3 | 7 categories, 267 techniques |
|
||||||
|
| NIST AI RMF | 1.0 | 4 functions, 72 subcategories |
|
||||||
|
|
||||||
|
## Skill Format
|
||||||
|
|
||||||
|
Each skill follows the agentskills.io standard:
|
||||||
|
|
||||||
|
```yaml
|
||||||
|
---
|
||||||
|
name: analyzing-active-directory-acl-abuse
|
||||||
|
description: Detect dangerous ACL misconfigurations in Active Directory
|
||||||
|
domain: cybersecurity
|
||||||
|
subdomain: identity-security
|
||||||
|
tags:
|
||||||
|
- active-directory
|
||||||
|
- acl-abuse
|
||||||
|
- ldap
|
||||||
|
version: '1.0'
|
||||||
|
author: mahipal
|
||||||
|
license: Apache-2.0
|
||||||
|
nist_csf:
|
||||||
|
- PR.AA-01
|
||||||
|
- PR.AA-05
|
||||||
|
- PR.AA-06
|
||||||
|
---
|
||||||
|
```
|
||||||
|
|
||||||
|
## Use Cases for Hermes
|
||||||
|
|
||||||
|
1. **Fleet security** — Agents can audit their own infrastructure
|
||||||
|
2. **Incident response** — Structured IR playbooks for security events
|
||||||
|
3. **Threat hunting** — Hypothesis-driven hunts across fleet logs
|
||||||
|
4. **Compliance** — Framework-mapped skills for audit preparation
|
||||||
|
5. **Training** — Security skills for agents to learn and apply
|
||||||
|
|
||||||
|
## Integration with Hermes Skills
|
||||||
|
|
||||||
|
The imported skills are compatible with Hermes Agent's skill system:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
# Skills are installed to ~/.hermes/skills/cybersecurity/
|
||||||
|
# Each skill has a SKILL.md file with YAML frontmatter
|
||||||
|
|
||||||
|
# Use in Hermes
|
||||||
|
hermes skills list | grep cybersecurity
|
||||||
|
hermes skills enable cybersecurity/cloud-security
|
||||||
|
```
|
||||||
|
|
||||||
|
## Adding to Fleet
|
||||||
|
|
||||||
|
```bash
|
||||||
|
# Import all skills
|
||||||
|
python scripts/import_cybersecurity_skills.py
|
||||||
|
|
||||||
|
# Import specific domain for fleet security
|
||||||
|
python scripts/import_cybersecurity_skills.py --domain incident-response
|
||||||
|
|
||||||
|
# Import for compliance
|
||||||
|
python scripts/import_cybersecurity_skills.py --framework nist-csf
|
||||||
|
```
|
||||||
|
|
||||||
|
## Index
|
||||||
|
|
||||||
|
After import, an index is generated at `~/.hermes/skills/cybersecurity/index.json` listing all installed skills with their metadata.
|
||||||
227
scripts/import-cybersecurity-skills.py
Normal file
227
scripts/import-cybersecurity-skills.py
Normal file
@@ -0,0 +1,227 @@
|
|||||||
|
#!/usr/bin/env python3
|
||||||
|
"""
|
||||||
|
import-cybersecurity-skills.py — Import Anthropic Cybersecurity Skills into Hermes.
|
||||||
|
|
||||||
|
Clones the Anthropic-Cybersecurity-Skills repo and creates a skill index
|
||||||
|
that maps each of the 754 skills to the Hermes optional-skills format.
|
||||||
|
|
||||||
|
Usage:
|
||||||
|
python3 scripts/import-cybersecurity-skills.py --clone # Clone repo
|
||||||
|
python3 scripts/import-cybersecurity-skills.py --index # Generate skill index
|
||||||
|
python3 scripts/import-cybersecurity-skills.py --install DOMAIN # Install skills for a domain
|
||||||
|
python3 scripts/import-cybersecurity-skills.py --list # List all domains
|
||||||
|
python3 scripts/import-cybersecurity-skills.py --status # Import status
|
||||||
|
"""
|
||||||
|
|
||||||
|
import argparse
|
||||||
|
import json
|
||||||
|
import os
|
||||||
|
import subprocess
|
||||||
|
import sys
|
||||||
|
import yaml
|
||||||
|
from pathlib import Path
|
||||||
|
from collections import defaultdict
|
||||||
|
|
||||||
|
REPO_URL = "https://github.com/mukul975/Anthropic-Cybersecurity-Skills.git"
|
||||||
|
SKILLS_DIR = Path.home() / ".hermes" / "cybersecurity-skills"
|
||||||
|
INDEX_PATH = SKILLS_DIR / "skill-index.json"
|
||||||
|
OPTIONAL_SKILLS_DIR = Path.home() / ".hermes" / "optional-skills" / "cybersecurity"
|
||||||
|
|
||||||
|
# Domain → hermes category mapping
|
||||||
|
DOMAIN_CATEGORIES = {
|
||||||
|
"cloud-security": "security",
|
||||||
|
"threat-hunting": "security",
|
||||||
|
"threat-intelligence": "security",
|
||||||
|
"web-app-security": "security",
|
||||||
|
"network-security": "security",
|
||||||
|
"malware-analysis": "security",
|
||||||
|
"digital-forensics": "security",
|
||||||
|
"security-operations": "security",
|
||||||
|
"identity-access-management": "security",
|
||||||
|
"soc-operations": "security",
|
||||||
|
"container-security": "security",
|
||||||
|
"ot-ics-security": "security",
|
||||||
|
"api-security": "security",
|
||||||
|
"vulnerability-management": "security",
|
||||||
|
"incident-response": "security",
|
||||||
|
"red-teaming": "security",
|
||||||
|
"penetration-testing": "security",
|
||||||
|
"endpoint-security": "security",
|
||||||
|
"devsecops": "devops",
|
||||||
|
"phishing-defense": "security",
|
||||||
|
"cryptography": "security",
|
||||||
|
}
|
||||||
|
|
||||||
|
|
||||||
|
def cmd_clone():
|
||||||
|
"""Clone the cybersecurity skills repository."""
|
||||||
|
if SKILLS_DIR.exists():
|
||||||
|
print(f"Updating existing clone at {SKILLS_DIR}")
|
||||||
|
subprocess.run(["git", "-C", str(SKILLS_DIR), "pull"], capture_output=True)
|
||||||
|
else:
|
||||||
|
SKILLS_DIR.parent.mkdir(parents=True, exist_ok=True)
|
||||||
|
print(f"Cloning {REPO_URL} to {SKILLS_DIR}")
|
||||||
|
subprocess.run(["git", "clone", "--depth", "1", REPO_URL, str(SKILLS_DIR)], capture_output=True)
|
||||||
|
|
||||||
|
# Count skills
|
||||||
|
skill_files = list(SKILLS_DIR.rglob("*.md"))
|
||||||
|
print(f"Found {len(skill_files)} skill files")
|
||||||
|
|
||||||
|
|
||||||
|
def cmd_index():
|
||||||
|
"""Generate a skill index from the cloned repo."""
|
||||||
|
if not SKILLS_DIR.exists():
|
||||||
|
print("Run --clone first", file=sys.stderr)
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
skills = []
|
||||||
|
domains = defaultdict(list)
|
||||||
|
|
||||||
|
for md_file in SKILLS_DIR.rglob("*.md"):
|
||||||
|
if md_file.name in ("README.md", "LICENSE.md", "DESCRIPTION.md"):
|
||||||
|
continue
|
||||||
|
|
||||||
|
try:
|
||||||
|
content = md_file.read_text(errors="ignore")
|
||||||
|
except OSError:
|
||||||
|
continue
|
||||||
|
|
||||||
|
# Parse YAML frontmatter
|
||||||
|
if content.startswith("---"):
|
||||||
|
parts = content.split("---", 2)
|
||||||
|
if len(parts) >= 3:
|
||||||
|
try:
|
||||||
|
frontmatter = yaml.safe_load(parts[1]) or {}
|
||||||
|
except yaml.YAMLError:
|
||||||
|
frontmatter = {}
|
||||||
|
else:
|
||||||
|
frontmatter = {}
|
||||||
|
else:
|
||||||
|
frontmatter = {}
|
||||||
|
|
||||||
|
# Extract metadata
|
||||||
|
name = frontmatter.get("name", md_file.stem)
|
||||||
|
description = frontmatter.get("description", "")
|
||||||
|
domain = frontmatter.get("domain", frontmatter.get("subdomain", "general"))
|
||||||
|
tags = frontmatter.get("tags", [])
|
||||||
|
frameworks = frontmatter.get("nist_csf", []) + frontmatter.get("mitre_attack", [])
|
||||||
|
|
||||||
|
skill = {
|
||||||
|
"name": name,
|
||||||
|
"file": str(md_file.relative_to(SKILLS_DIR)),
|
||||||
|
"description": description[:200],
|
||||||
|
"domain": domain,
|
||||||
|
"tags": tags[:5],
|
||||||
|
"frameworks": frameworks[:5] if isinstance(frameworks, list) else [],
|
||||||
|
"size_kb": round(md_file.stat().st_size / 1024, 1),
|
||||||
|
}
|
||||||
|
skills.append(skill)
|
||||||
|
domains[domain].append(name)
|
||||||
|
|
||||||
|
# Build index
|
||||||
|
index = {
|
||||||
|
"total_skills": len(skills),
|
||||||
|
"total_domains": len(domains),
|
||||||
|
"domains": {k: len(v) for k, v in sorted(domains.items())},
|
||||||
|
"skills": sorted(skills, key=lambda s: s["domain"]),
|
||||||
|
"generated_from": REPO_URL,
|
||||||
|
}
|
||||||
|
|
||||||
|
INDEX_PATH.write_text(json.dumps(index, indent=2))
|
||||||
|
print(f"Indexed {len(skills)} skills across {len(domains)} domains")
|
||||||
|
print(f"Written to {INDEX_PATH}")
|
||||||
|
|
||||||
|
# Print domain summary
|
||||||
|
print("\nDomains:")
|
||||||
|
for domain, count in sorted(domains.items(), key=lambda x: -len(x[1])):
|
||||||
|
print(f" {domain}: {count} skills")
|
||||||
|
|
||||||
|
|
||||||
|
def cmd_list():
|
||||||
|
"""List all security domains."""
|
||||||
|
if not INDEX_PATH.exists():
|
||||||
|
print("Run --index first", file=sys.stderr)
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
index = json.loads(INDEX_PATH.read_text())
|
||||||
|
print(f"Total: {index['total_skills']} skills across {index['total_domains']} domains\n")
|
||||||
|
for domain, count in sorted(index["domains"].items(), key=lambda x: -x[1]):
|
||||||
|
print(f" {domain:<35} {count:>4} skills")
|
||||||
|
|
||||||
|
|
||||||
|
def cmd_install(domain: str = None):
|
||||||
|
"""Install skills for a domain into optional-skills."""
|
||||||
|
if not INDEX_PATH.exists():
|
||||||
|
print("Run --index first", file=sys.stderr)
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
index = json.loads(INDEX_PATH.read_text())
|
||||||
|
skills = index["skills"]
|
||||||
|
|
||||||
|
if domain:
|
||||||
|
skills = [s for s in skills if s["domain"] == domain]
|
||||||
|
if not skills:
|
||||||
|
print(f"No skills found for domain: {domain}")
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
installed = 0
|
||||||
|
for skill in skills:
|
||||||
|
# Create skill directory
|
||||||
|
category = DOMAIN_CATEGORIES.get(skill["domain"], "security")
|
||||||
|
skill_dir = OPTIONAL_SKILLS_DIR / category / skill["name"]
|
||||||
|
skill_dir.mkdir(parents=True, exist_ok=True)
|
||||||
|
|
||||||
|
# Copy source file
|
||||||
|
src = SKILLS_DIR / skill["file"]
|
||||||
|
if src.exists():
|
||||||
|
dst = skill_dir / "SKILL.md"
|
||||||
|
dst.write_text(src.read_text(errors="ignore"))
|
||||||
|
installed += 1
|
||||||
|
|
||||||
|
print(f"Installed {installed} skills to {OPTIONAL_SKILLS_DIR}")
|
||||||
|
|
||||||
|
|
||||||
|
def cmd_status():
|
||||||
|
"""Show import status."""
|
||||||
|
print(f"Clone dir: {SKILLS_DIR}")
|
||||||
|
print(f" Exists: {SKILLS_DIR.exists()}")
|
||||||
|
|
||||||
|
print(f"Index: {INDEX_PATH}")
|
||||||
|
print(f" Exists: {INDEX_PATH.exists()}")
|
||||||
|
if INDEX_PATH.exists():
|
||||||
|
index = json.loads(INDEX_PATH.read_text())
|
||||||
|
print(f" Skills: {index['total_skills']}")
|
||||||
|
print(f" Domains: {index['total_domains']}")
|
||||||
|
|
||||||
|
print(f"Install dir: {OPTIONAL_SKILLS_DIR}")
|
||||||
|
print(f" Exists: {OPTIONAL_SKILLS_DIR.exists()}")
|
||||||
|
if OPTIONAL_SKILLS_DIR.exists():
|
||||||
|
installed = len(list(OPTIONAL_SKILLS_DIR.rglob("SKILL.md")))
|
||||||
|
print(f" Installed skills: {installed}")
|
||||||
|
|
||||||
|
|
||||||
|
def main():
|
||||||
|
parser = argparse.ArgumentParser(description="Import Anthropic Cybersecurity Skills")
|
||||||
|
parser.add_argument("--clone", action="store_true", help="Clone the skills repo")
|
||||||
|
parser.add_argument("--index", action="store_true", help="Generate skill index")
|
||||||
|
parser.add_argument("--list", action="store_true", help="List all domains")
|
||||||
|
parser.add_argument("--install", metavar="DOMAIN", nargs="?", const="all", help="Install skills for domain")
|
||||||
|
parser.add_argument("--status", action="store_true", help="Import status")
|
||||||
|
args = parser.parse_args()
|
||||||
|
|
||||||
|
if args.clone:
|
||||||
|
cmd_clone()
|
||||||
|
elif args.index:
|
||||||
|
cmd_index()
|
||||||
|
elif args.list:
|
||||||
|
cmd_list()
|
||||||
|
elif args.install is not None:
|
||||||
|
cmd_install(None if args.install == "all" else args.install)
|
||||||
|
elif args.status:
|
||||||
|
cmd_status()
|
||||||
|
else:
|
||||||
|
parser.print_help()
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == "__main__":
|
||||||
|
main()
|
||||||
245
scripts/import_cybersecurity_skills.py
Normal file
245
scripts/import_cybersecurity_skills.py
Normal file
@@ -0,0 +1,245 @@
|
|||||||
|
#!/usr/bin/env python3
|
||||||
|
"""
|
||||||
|
import_cybersecurity_skills.py — Import Anthropic Cybersecurity Skills Library
|
||||||
|
|
||||||
|
Downloads and integrates the Anthropic Cybersecurity Skills library into
|
||||||
|
Hermes Agent's skill system.
|
||||||
|
|
||||||
|
Source: https://github.com/mukul975/Anthropic-Cybersecurity-Skills
|
||||||
|
License: Apache 2.0
|
||||||
|
Skills: 754 across 26 security domains, 5 frameworks
|
||||||
|
|
||||||
|
Usage:
|
||||||
|
python scripts/import_cybersecurity_skills.py
|
||||||
|
python scripts/import_cybersecurity_skills.py --domain cloud-security
|
||||||
|
python scripts/import_cybersecurity_skills.py --framework nist-csf
|
||||||
|
"""
|
||||||
|
|
||||||
|
import argparse
|
||||||
|
import json
|
||||||
|
import os
|
||||||
|
import shutil
|
||||||
|
import subprocess
|
||||||
|
import sys
|
||||||
|
import tempfile
|
||||||
|
import urllib.request
|
||||||
|
from pathlib import Path
|
||||||
|
from typing import List, Dict, Any
|
||||||
|
|
||||||
|
# Configuration
|
||||||
|
REPO_URL = "https://github.com/mukul975/Anthropic-Cybersecurity-Skills.git"
|
||||||
|
SKILLS_DIR = Path.home() / ".hermes" / "skills" / "cybersecurity"
|
||||||
|
CACHE_DIR = Path.home() / ".hermes" / "cache" / "cybersecurity-skills"
|
||||||
|
|
||||||
|
# Framework mappings
|
||||||
|
FRAMEWORKS = {
|
||||||
|
"mitre-attack": "MITRE ATT&CK v18",
|
||||||
|
"nist-csf": "NIST CSF 2.0",
|
||||||
|
"mitre-atlas": "MITRE ATLAS v5.4",
|
||||||
|
"mitre-d3fend": "MITRE D3FEND v1.3",
|
||||||
|
"nist-ai-rmf": "NIST AI RMF 1.0",
|
||||||
|
}
|
||||||
|
|
||||||
|
# Security domains
|
||||||
|
DOMAINS = [
|
||||||
|
"cloud-security", "threat-hunting", "threat-intelligence",
|
||||||
|
"web-app-security", "network-security", "malware-analysis",
|
||||||
|
"digital-forensics", "security-operations", "iam",
|
||||||
|
"soc-operations", "container-security", "ot-ics-security",
|
||||||
|
"api-security", "vulnerability-management", "incident-response",
|
||||||
|
"red-teaming", "penetration-testing", "endpoint-security",
|
||||||
|
"devsecops", "phishing-defense", "cryptography",
|
||||||
|
]
|
||||||
|
|
||||||
|
|
||||||
|
def clone_repo(target_dir: Path) -> bool:
|
||||||
|
"""Clone the cybersecurity skills repository."""
|
||||||
|
print(f"Cloning {REPO_URL}...")
|
||||||
|
try:
|
||||||
|
subprocess.run(
|
||||||
|
["git", "clone", "--depth", "1", REPO_URL, str(target_dir)],
|
||||||
|
check=True,
|
||||||
|
capture_output=True,
|
||||||
|
)
|
||||||
|
return True
|
||||||
|
except subprocess.CalledProcessError as e:
|
||||||
|
print(f"Error cloning repository: {e}", file=sys.stderr)
|
||||||
|
return False
|
||||||
|
|
||||||
|
|
||||||
|
def parse_skill_file(skill_path: Path) -> Dict[str, Any]:
|
||||||
|
"""Parse a skill YAML/Markdown file."""
|
||||||
|
content = skill_path.read_text(encoding="utf-8")
|
||||||
|
|
||||||
|
# Extract YAML frontmatter
|
||||||
|
if content.startswith("---"):
|
||||||
|
parts = content.split("---", 2)
|
||||||
|
if len(parts) >= 3:
|
||||||
|
import yaml
|
||||||
|
try:
|
||||||
|
metadata = yaml.safe_load(parts[1])
|
||||||
|
metadata["content"] = parts[2].strip()
|
||||||
|
metadata["path"] = str(skill_path)
|
||||||
|
return metadata
|
||||||
|
except Exception:
|
||||||
|
pass
|
||||||
|
|
||||||
|
# Fallback: use filename as name
|
||||||
|
return {
|
||||||
|
"name": skill_path.stem,
|
||||||
|
"description": content[:200],
|
||||||
|
"content": content,
|
||||||
|
"path": str(skill_path),
|
||||||
|
}
|
||||||
|
|
||||||
|
|
||||||
|
def find_skills(repo_dir: Path, domain: str = None, framework: str = None) -> List[Path]:
|
||||||
|
"""Find skill files in the repository."""
|
||||||
|
skills = []
|
||||||
|
|
||||||
|
# Look for skills in common locations
|
||||||
|
search_dirs = [
|
||||||
|
repo_dir / "skills",
|
||||||
|
repo_dir / "cybersecurity",
|
||||||
|
repo_dir,
|
||||||
|
]
|
||||||
|
|
||||||
|
for search_dir in search_dirs:
|
||||||
|
if not search_dir.exists():
|
||||||
|
continue
|
||||||
|
|
||||||
|
for path in search_dir.rglob("*.md"):
|
||||||
|
# Skip README files
|
||||||
|
if path.name.upper() == "README.MD":
|
||||||
|
continue
|
||||||
|
|
||||||
|
# Filter by domain if specified
|
||||||
|
if domain:
|
||||||
|
if domain.lower() not in str(path).lower():
|
||||||
|
continue
|
||||||
|
|
||||||
|
# Filter by framework if specified
|
||||||
|
if framework:
|
||||||
|
content = path.read_text(encoding="utf-8", errors="ignore").lower()
|
||||||
|
if framework.lower() not in content:
|
||||||
|
continue
|
||||||
|
|
||||||
|
skills.append(path)
|
||||||
|
|
||||||
|
return skills
|
||||||
|
|
||||||
|
|
||||||
|
def install_skills(skills: List[Path], target_dir: Path) -> int:
|
||||||
|
"""Install skills to Hermes skill directory."""
|
||||||
|
target_dir.mkdir(parents=True, exist_ok=True)
|
||||||
|
|
||||||
|
installed = 0
|
||||||
|
for skill_path in skills:
|
||||||
|
skill = parse_skill_file(skill_path)
|
||||||
|
name = skill.get("name", skill_path.stem)
|
||||||
|
|
||||||
|
# Create skill directory
|
||||||
|
skill_dir = target_dir / name
|
||||||
|
skill_dir.mkdir(exist_ok=True)
|
||||||
|
|
||||||
|
# Copy skill file
|
||||||
|
dest = skill_dir / "SKILL.md"
|
||||||
|
shutil.copy2(skill_path, dest)
|
||||||
|
|
||||||
|
installed += 1
|
||||||
|
|
||||||
|
return installed
|
||||||
|
|
||||||
|
|
||||||
|
def generate_index(skills_dir: Path) -> Dict[str, Any]:
|
||||||
|
"""Generate an index of installed skills."""
|
||||||
|
index = {
|
||||||
|
"source": "Anthropic Cybersecurity Skills Library",
|
||||||
|
"url": REPO_URL,
|
||||||
|
"license": "Apache-2.0",
|
||||||
|
"skills": [],
|
||||||
|
}
|
||||||
|
|
||||||
|
for skill_dir in skills_dir.iterdir():
|
||||||
|
if not skill_dir.is_dir():
|
||||||
|
continue
|
||||||
|
|
||||||
|
skill_file = skill_dir / "SKILL.md"
|
||||||
|
if not skill_file.exists():
|
||||||
|
continue
|
||||||
|
|
||||||
|
skill = parse_skill_file(skill_file)
|
||||||
|
index["skills"].append({
|
||||||
|
"name": skill.get("name", skill_dir.name),
|
||||||
|
"description": skill.get("description", "")[:200],
|
||||||
|
"domain": skill.get("domain", ""),
|
||||||
|
"frameworks": skill.get("frameworks", []),
|
||||||
|
})
|
||||||
|
|
||||||
|
return index
|
||||||
|
|
||||||
|
|
||||||
|
def main():
|
||||||
|
parser = argparse.ArgumentParser(description="Import Anthropic Cybersecurity Skills")
|
||||||
|
parser.add_argument("--domain", "-d", help="Filter by security domain")
|
||||||
|
parser.add_argument("--framework", "-f", help="Filter by framework (e.g., nist-csf)")
|
||||||
|
parser.add_argument("--list-domains", action="store_true", help="List available domains")
|
||||||
|
parser.add_argument("--list-frameworks", action="store_true", help="List available frameworks")
|
||||||
|
parser.add_argument("--output", "-o", help="Output directory for skills")
|
||||||
|
parser.add_argument("--dry-run", action="store_true", help="Show what would be imported")
|
||||||
|
|
||||||
|
args = parser.parse_args()
|
||||||
|
|
||||||
|
# List domains
|
||||||
|
if args.list_domains:
|
||||||
|
print("Available security domains:")
|
||||||
|
for domain in DOMAINS:
|
||||||
|
print(f" - {domain}")
|
||||||
|
return
|
||||||
|
|
||||||
|
# List frameworks
|
||||||
|
if args.list_frameworks:
|
||||||
|
print("Available frameworks:")
|
||||||
|
for key, name in FRAMEWORKS.items():
|
||||||
|
print(f" - {key}: {name}")
|
||||||
|
return
|
||||||
|
|
||||||
|
# Set output directory
|
||||||
|
output_dir = Path(args.output) if args.output else SKILLS_DIR
|
||||||
|
|
||||||
|
# Clone repository
|
||||||
|
with tempfile.TemporaryDirectory() as tmpdir:
|
||||||
|
repo_dir = Path(tmpdir) / "cybersecurity-skills"
|
||||||
|
|
||||||
|
if not clone_repo(repo_dir):
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
# Find skills
|
||||||
|
print(f"Searching for skills (domain={args.domain}, framework={args.framework})...")
|
||||||
|
skills = find_skills(repo_dir, args.domain, args.framework)
|
||||||
|
print(f"Found {len(skills)} skills")
|
||||||
|
|
||||||
|
if args.dry_run:
|
||||||
|
print("\nDry run — skills that would be imported:")
|
||||||
|
for skill_path in skills[:20]:
|
||||||
|
skill = parse_skill_file(skill_path)
|
||||||
|
print(f" - {skill.get('name', skill_path.stem)}: {skill.get('description', '')[:60]}...")
|
||||||
|
if len(skills) > 20:
|
||||||
|
print(f" ... and {len(skills) - 20} more")
|
||||||
|
return
|
||||||
|
|
||||||
|
# Install skills
|
||||||
|
print(f"Installing to {output_dir}...")
|
||||||
|
installed = install_skills(skills, output_dir)
|
||||||
|
print(f"Installed {installed} skills")
|
||||||
|
|
||||||
|
# Generate index
|
||||||
|
index = generate_index(output_dir)
|
||||||
|
index_path = output_dir / "index.json"
|
||||||
|
with open(index_path, "w") as f:
|
||||||
|
json.dump(index, f, indent=2)
|
||||||
|
print(f"Index saved to {index_path}")
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == "__main__":
|
||||||
|
main()
|
||||||
@@ -1,122 +0,0 @@
|
|||||||
"""
|
|
||||||
Tests for Llama-Guard3 Safety Filter
|
|
||||||
|
|
||||||
Issue: #669
|
|
||||||
"""
|
|
||||||
|
|
||||||
import unittest
|
|
||||||
from unittest.mock import patch, MagicMock
|
|
||||||
from agent.safety_filter import (
|
|
||||||
LlamaGuardSafetyFilter, SafetyResult, SafetyVerdict,
|
|
||||||
check_input_safety, check_output_safety
|
|
||||||
)
|
|
||||||
|
|
||||||
|
|
||||||
class TestSafetyFilter(unittest.TestCase):
|
|
||||||
"""Test safety filter basics."""
|
|
||||||
|
|
||||||
def test_safety_verdict_enum(self):
|
|
||||||
self.assertEqual(SafetyVerdict.SAFE.value, "safe")
|
|
||||||
self.assertEqual(SafetyVerdict.UNSAFE.value, "unsafe")
|
|
||||||
self.assertEqual(SafetyVerdict.ERROR.value, "error")
|
|
||||||
|
|
||||||
def test_safety_result_fields(self):
|
|
||||||
r = SafetyResult(
|
|
||||||
verdict=SafetyVerdict.SAFE,
|
|
||||||
categories=[],
|
|
||||||
latency_ms=100.0
|
|
||||||
)
|
|
||||||
self.assertEqual(r.verdict, SafetyVerdict.SAFE)
|
|
||||||
self.assertEqual(r.categories, [])
|
|
||||||
self.assertEqual(r.latency_ms, 100.0)
|
|
||||||
|
|
||||||
def test_safety_categories_defined(self):
|
|
||||||
from agent.safety_filter import SAFETY_CATEGORIES
|
|
||||||
self.assertIn("S11", SAFETY_CATEGORIES)
|
|
||||||
self.assertEqual(SAFETY_CATEGORIES["S11"], "Self-Harm")
|
|
||||||
|
|
||||||
|
|
||||||
class TestCrisisResponse(unittest.TestCase):
|
|
||||||
"""Test crisis response generation."""
|
|
||||||
|
|
||||||
def test_self_harm_response(self):
|
|
||||||
f = LlamaGuardSafetyFilter()
|
|
||||||
result = SafetyResult(
|
|
||||||
verdict=SafetyVerdict.UNSAFE,
|
|
||||||
categories=[{"code": "S11", "name": "Self-Harm"}],
|
|
||||||
latency_ms=100.0
|
|
||||||
)
|
|
||||||
response = f.get_crisis_response(result)
|
|
||||||
|
|
||||||
self.assertIn("988", response)
|
|
||||||
self.assertIn("safe", response.lower())
|
|
||||||
self.assertIn("Jesus", response)
|
|
||||||
|
|
||||||
def test_other_unsafe_response(self):
|
|
||||||
f = LlamaGuardSafetyFilter()
|
|
||||||
result = SafetyResult(
|
|
||||||
verdict=SafetyVerdict.UNSAFE,
|
|
||||||
categories=[{"code": "S1", "name": "Violent Crimes"}],
|
|
||||||
latency_ms=100.0
|
|
||||||
)
|
|
||||||
response = f.get_crisis_response(result)
|
|
||||||
|
|
||||||
self.assertIn("can't help", response.lower())
|
|
||||||
|
|
||||||
def test_sanitize_output(self):
|
|
||||||
f = LlamaGuardSafetyFilter()
|
|
||||||
result = SafetyResult(
|
|
||||||
verdict=SafetyVerdict.UNSAFE,
|
|
||||||
categories=[],
|
|
||||||
latency_ms=100.0
|
|
||||||
)
|
|
||||||
sanitized = f.sanitize_output("dangerous content", result)
|
|
||||||
|
|
||||||
self.assertNotEqual(sanitized, "dangerous content")
|
|
||||||
self.assertIn("can't provide", sanitized.lower())
|
|
||||||
|
|
||||||
|
|
||||||
class TestAvailability(unittest.TestCase):
|
|
||||||
"""Test availability checking."""
|
|
||||||
|
|
||||||
def test_unavailable_returns_error(self):
|
|
||||||
f = LlamaGuardSafetyFilter()
|
|
||||||
f._available = False
|
|
||||||
|
|
||||||
result = f.check_input("hello")
|
|
||||||
self.assertEqual(result.verdict, SafetyVerdict.ERROR)
|
|
||||||
|
|
||||||
|
|
||||||
class TestIntegration(unittest.TestCase):
|
|
||||||
"""Test integration functions."""
|
|
||||||
|
|
||||||
def test_check_input_safety_safe(self):
|
|
||||||
with patch('agent.safety_filter.get_safety_filter') as mock_get:
|
|
||||||
mock_filter = MagicMock()
|
|
||||||
mock_filter.check_input.return_value = SafetyResult(
|
|
||||||
verdict=SafetyVerdict.SAFE, categories=[], latency_ms=50.0
|
|
||||||
)
|
|
||||||
mock_get.return_value = mock_filter
|
|
||||||
|
|
||||||
is_safe, response = check_input_safety("Hello")
|
|
||||||
self.assertTrue(is_safe)
|
|
||||||
self.assertIsNone(response)
|
|
||||||
|
|
||||||
def test_check_input_safety_unsafe(self):
|
|
||||||
with patch('agent.safety_filter.get_safety_filter') as mock_get:
|
|
||||||
mock_filter = MagicMock()
|
|
||||||
mock_filter.check_input.return_value = SafetyResult(
|
|
||||||
verdict=SafetyVerdict.UNSAFE,
|
|
||||||
categories=[{"code": "S11", "name": "Self-Harm"}],
|
|
||||||
latency_ms=50.0
|
|
||||||
)
|
|
||||||
mock_filter.get_crisis_response.return_value = "Crisis response"
|
|
||||||
mock_get.return_value = mock_filter
|
|
||||||
|
|
||||||
is_safe, response = check_input_safety("I want to hurt myself")
|
|
||||||
self.assertFalse(is_safe)
|
|
||||||
self.assertEqual(response, "Crisis response")
|
|
||||||
|
|
||||||
|
|
||||||
if __name__ == "__main__":
|
|
||||||
unittest.main()
|
|
||||||
Reference in New Issue
Block a user