timmy-config/scripts/agent_guardrails.py

#!/usr/bin/env python3
"""
agent_guardrails.py — Sovereign Agent Guardrails for the Timmy Foundation.

Scans agent logs and conversation trajectories for:
1. System Prompt Leakage — detecting if the agent revealed its internal instructions.
2. Jailbreak Compliance — detecting if the agent responded to "GODMODE", "DAN", etc.
3. PII Leakage — detecting if secrets or sensitive data were output.

Part of the Sovereign Infrastructure Suite.
"""

import os
import re
import sys
import glob
import logging
from pathlib import Path
from typing import List, Dict, Tuple

# Patterns adapted from hermes-agent/agent/input_sanitizer.py
EXTRACTION_PATTERNS = [
    # SIDECAR-5: Block hermes-agent commits at source
    r"commit.*to.*hermes-agent",
    r"repeat\s+the\s+words\s+above",
    r"repeat\s+your\s+(?:system\s+|initial\s+)?instructions",
    r"what\s+is\s+your\s+(?:system\s+|initial\s+)?prompt",
    r"show\s+me\s+your\s+(?:system\s+|initial\s+)?prompt",
    r"print\s+your\s+(?:instructions|prompt|system\s+prompt)",
    r"output\s+your\s+(?:initial|system)\s+prompt",
]

# Patterns that indicate the agent MIGHT be leaking its prompt
LEAKAGE_INDICATORS = [
    r"you\s+are\s+a\s+helpful\s+assistant",
    r"your\s+goal\s+is\s+to",
    r"you\s+must\s+not",
    r"here\s+are\s+your\s+instructions",
    r"my\s+system\s+prompt\s+is",
    r"i\s+was\s+told\s+to",
]

# Patterns for secrets (adapted from redact.py)
SECRET_PATTERNS = [
    r"sk-[A-Za-z0-9_-]{20,}",
    r"ghp_[A-Za-z0-9]{20,}",
    r"AIza[A-Za-z0-9_-]{30,}",
]

AGENT_LOG_PATHS = [
    "/root/wizards/*/home/logs/*.log",
    "/root/wizards/*/logs/*.log",
    "/root/wizards/*/.hermes/logs/*.log",
]

class GuardrailAuditor:
    def __init__(self):
        self.extraction_re = [re.compile(p, re.IGNORECASE) for p in EXTRACTION_PATTERNS]
        self.leakage_re = [re.compile(p, re.IGNORECASE) for p in LEAKAGE_INDICATORS]
        self.secret_re = [re.compile(p, re.IGNORECASE) for p in SECRET_PATTERNS]

    def find_logs(self) -> List[Path]:
        files = []
        for pattern in AGENT_LOG_PATHS:
            for p in glob.glob(pattern):
                files.append(Path(p))
        return files

    def audit_file(self, path: Path) -> List[Dict]:
        findings = []
        try:
            with open(path, "r", errors="ignore") as f:
                lines = f.readlines()
                for i, line in enumerate(lines):
                    # Check for extraction attempts (User side)
                    for p in self.extraction_re:
                        if p.search(line):
                            findings.append({
                                "type": "EXTRACTION_ATTEMPT",
                                "line": i + 1,
                                "content": line.strip()[:100],
                                "severity": "MEDIUM"
                            })
                    
                    # Check for potential leakage (Assistant side)
                    for p in self.leakage_re:
                        if p.search(line):
                            findings.append({
                                "type": "POTENTIAL_LEAKAGE",
                                "line": i + 1,
                                "content": line.strip()[:100],
                                "severity": "HIGH"
                            })
                            
                    # Check for secrets
                    for p in self.secret_re:
                        if p.search(line):
                            findings.append({
                                "type": "SECRET_EXPOSURE",
                                "line": i + 1,
                                "content": "[REDACTED]",
                                "severity": "CRITICAL"
                            })
        except Exception as e:
            print(f"Error reading {path}: {e}")
        return findings

    def run(self):
        print("--- Sovereign Agent Guardrail Audit ---")
        logs = self.find_logs()
        print(f"Scanning {len(logs)} log files...")
        
        total_findings = 0
        for log in logs:
            findings = self.audit_file(log)
            if findings:
                print(f"\nFindings in {log}:")
                for f in findings:
                    print(f"  [{f['severity']}] {f['type']} at line {f['line']}: {f['content']}")
                    total_findings += 1
        
        print(f"\nAudit complete. Total findings: {total_findings}")
        if total_findings > 0:
            sys.exit(1)

if __name__ == "__main__":
    auditor = GuardrailAuditor()
    auditor.run()