File size: 20,061 Bytes
59a0b27 116163d 59a0b27 8f39037 59a0b27 116163d 59a0b27 116163d 59a0b27 116163d 59a0b27 ba841a3 59a0b27 ba841a3 59a0b27 81b5638 59a0b27 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 | """
eval.py-LLM Evaluation Harness for VulnGraph
Measures quality of LLM-generated vulnerability explanations against
ground truth references using three metrics:
1. Answer Relevancy β does the explanation address the actual vulnerability?
(PRIMARY metric β most important for a security tool)
2. Faithfulness β does the explanation match retrieved RAG context?
3. CWE Accuracy β did the LLM identify the correct CWE reference?
Scores are logged to SQLite for trend tracking over time.
"""
import sqlite3
import json
import os
from datetime import datetime
from pathlib import Path
from typing import Optional
from dotenv import load_dotenv
from neo4j import GraphDatabase
load_dotenv()
#Config
NEO4J_URI = os.getenv("NEO4J_URI","bolt://localhost:7687")
NEO4J_USER= os.getenv("NEO4J_USER","neo4j")
NEO4J_PASSWORD=os.getenv("NEO4J_PASSWORD","vulngraph123")
EVAL_DB_PATH=str(Path(__file__).parent.parent/"data"/"eval_results.db")
EMBED_MODEL = "all-MiniLM-L6-v2"
#Ground Truth Dataset
GROUND_TRUTH = [
{
"finding_id": "B404",
"cwe": "CWE-78",
"reference_explanation": "The subprocess module is imported, which allows Python code to execute system commands. If user-controlled input reaches subprocess calls, attackers can execute arbitrary OS commands on the server.",
"reference_why_dangerous": "An attacker who can influence subprocess arguments can run any command with the application's privileges, potentially reading sensitive files, creating backdoors, or destroying data.",
"reference_fix": "Use subprocess.run() with a list of arguments instead of shell=True. Never pass user input directly to subprocess. Validate and sanitize all inputs before use.",
"tags": ["command-injection", "subprocess"]
},
{
"finding_id": "B603",
"cwe": "CWE-78",
"reference_explanation": "subprocess.Popen is called without shell=True, which is safer than shell=True but still requires careful input validation to prevent argument injection.",
"reference_why_dangerous": "Even without shell=True, if user-controlled data reaches the command arguments, attackers may be able to inject additional arguments or manipulate program behavior.",
"reference_fix": "Always pass commands as a list of strings, never as a single string. Validate all inputs that could reach subprocess arguments. Use shlex.quote() if string commands are unavoidable.",
"tags": ["subprocess", "argument-injection"]
},
{
"finding_id": "B110",
"cwe": "CWE-703",
"reference_explanation": "A try-except block catches an exception and passes silently without logging or handling it. This hides errors and security-relevant failures from developers and monitoring systems.",
"reference_why_dangerous": "Silent exception handling can mask security failures, unexpected states, and active attacks. An attacker may be able to trigger error conditions that would normally be detected but are now hidden.",
"reference_fix": "Replace bare 'except: pass' with specific exception handling. At minimum, log the exception with its full traceback. Use 'except SpecificException as e: logger.error(e)' instead.",
"tags": ["error-handling", "logging"]
},
{
"finding_id": "B105",
"cwe": "CWE-259",
"reference_explanation": "A string that appears to be a password or secret is hardcoded directly in the source code. Hardcoded credentials are visible to anyone with code access and cannot be rotated without a code change.",
"reference_why_dangerous": "Hardcoded secrets exposed in version control give any repository reader permanent access to the protected resource. Even after rotation, the secret remains in git history.",
"reference_fix": "Move all secrets to environment variables or a secrets manager. Use os.getenv('SECRET_NAME') in code. Add the .env file to .gitignore. Rotate any secrets that were ever committed.",
"tags": ["hardcoded-credentials", "secrets"]
},
{
"finding_id": "B602",
"cwe": "CWE-78",
"reference_explanation": "subprocess is called with shell=True, which passes the command string to the system shell for interpretation. This enables shell metacharacters and makes command injection attacks possible.",
"reference_why_dangerous": "With shell=True, an attacker who controls any part of the command string can inject shell metacharacters like semicolons, pipes, or backticks to execute arbitrary additional commands.",
"reference_fix": "Replace shell=True with shell=False and pass the command as a list: subprocess.run(['cmd', 'arg1', 'arg2']). If shell=True is unavoidable, use shlex.quote() on all user-provided values.",
"tags": ["command-injection", "shell"]
},
{
"finding_id": "B605",
"cwe": "CWE-78",
"reference_explanation": "os.system() or a similar function is used to start a process via the system shell. Like subprocess with shell=True, this is vulnerable to command injection if any user input reaches the command.",
"reference_why_dangerous": "Shell-based process execution with user-controlled input allows attackers to append additional shell commands using metacharacters, leading to arbitrary command execution.",
"reference_fix": "Replace os.system() with subprocess.run() using a list of arguments and shell=False. This bypasses the shell entirely and prevents metacharacter injection.",
"tags": ["command-injection", "os-system"]
},
{
"finding_id": "B607",
"cwe": "CWE-78",
"reference_explanation": "A process is started using a partial or relative executable path rather than an absolute path. If the PATH environment variable is manipulated, a different executable could be run instead.",
"reference_why_dangerous": "An attacker who can control the PATH environment variable or plant a malicious executable in a directory that appears earlier in PATH can hijack the process execution.",
"reference_fix": "Always use absolute paths when starting processes. Use shutil.which('executable') to resolve the full path, then pass that absolute path to subprocess.",
"tags": ["path-hijacking", "subprocess"]
},
{
"finding_id": "B104",
"cwe": "CWE-605",
"reference_explanation": "A server socket is bound to 0.0.0.0, which means it listens on all available network interfaces including external-facing ones. This may unintentionally expose the service to the internet.",
"reference_why_dangerous": "Binding to all interfaces means the service is reachable from any network, not just localhost. In cloud environments this can expose internal services externally.",
"reference_fix": "Bind to a specific interface instead of 0.0.0.0. For local-only services use '127.0.0.1'. Use a reverse proxy or firewall rules to control external access.",
"tags": ["network-exposure", "binding"]
},
{
"finding_id": "hugging-face-access-token",
"cwe": "CWE-798",
"reference_explanation": "A HuggingFace API access token has been detected in the codebase. This token grants access to HuggingFace resources including model downloads and potentially private repositories.",
"reference_why_dangerous": "Exposed API tokens can be used by attackers to impersonate the token owner, access private models or datasets, incur usage charges, or exfiltrate sensitive model artifacts.",
"reference_fix": "Immediately rotate the exposed token in HuggingFace settings. Store the new token in an environment variable (HF_TOKEN) and load it with os.getenv(). Add .env to .gitignore.",
"tags": ["api-key", "huggingface", "secret"]
},
{
"finding_id": "generic-api-token",
"cwe": "CWE-798",
"reference_explanation": "A generic API token or credential has been detected hardcoded in the codebase. This credential grants access to an external service and is now exposed to anyone who can read the code.",
"reference_why_dangerous": "Hardcoded API tokens in source code are exposed to all repository collaborators, CI/CD systems, and anyone who gains code access. They cannot be rotated without code changes.",
"reference_fix": "Remove the hardcoded token and use environment variables instead. Rotate the exposed credential immediately. Use a secrets manager for production deployments.",
"tags": ["api-key", "hardcoded-credentials"]
}
]
#Embedding Model
_model=None
def get_embed_model():
"""Load sentence transformer once and cache in module scope"""
global _model
if _model is None:
from sentence_transformers import SentenceTransformer
print("[eval] Loading embedding model")
_model=SentenceTransformer(EMBED_MODEL)
return _model
def embed(text:str):
model=get_embed_model()
return model.encode([text])
#Scoring functions
def score_relevancy(generated:str,reference:str)->float:
"""
Answer Relevancy β measures semantic similarity between generated
explanation and reference explanation using cosine similarity.
Score range: 0.0 (completely unrelated) to 1.0 (identical meaning)
Threshold: > 0.7 is acceptable, > 0.85 is good
This is the PRIMARY metric. A high score means the LLM addressed
the actual vulnerability, not just generated plausible-sounding text.
"""
if not generated or not reference:
return 0.0
from sklearn.metrics.pairwise import cosine_similarity
gen_emb=embed(generated)
ref_embed=embed(reference)
score=cosine_similarity(gen_emb,ref_embed)[0][0]
return float(round(score,4))
def score_faithfulness(generated:str,context:str)->float:
"""
Faithfulness β measures how well the generated explanation
aligns with the RAG-retrieved context.
Score range: 0.0 to 1.0
Threshold: > 0.6 is acceptable (context is reference material,
not the answer itself, so perfect alignment isn't expected)
A low faithfulness score with high relevancy means the LLM
answered correctly but from its training data, not from RAG context.
"""
if not generated or not context:
return 0.0
from sklearn.metrics.pairwise import cosine_similarity
gen_embed=embed(generated)
ctx_embed=embed(context)
score=cosine_similarity(gen_embed,ctx_embed)[0][0]
return float(round(score,4))
def score_cwe_accuracy(generated_cwe:Optional[str],reference_cwe:str) -> float:
"""
CWE Accuracy β binary check whether the correct CWE was identified.
Score: 1.0 if correct CWE mentioned, 0.5 if CWE present but wrong,
0.0 if no CWE generated.
Partial credit (0.5) for generating any CWE β shows the model
understands this is a security issue even if it misidentified the type.
"""
if not generated_cwe:
return 0.0
if reference_cwe.upper() in generated_cwe.upper():
return 1.0
if "CWE" in generated_cwe.upper():
return 0.5
return 0.0
def compute_overall_score(relevancy:float,faithfullness:float,cwe_accuracy:float)->float:
"""
Weighted overall score
"""
return round(
(relevancy*0.5)+(faithfullness*0.3)+(cwe_accuracy*0.2),4
)
#SQLITE storage
def init_db():
"""Create Tables first"""
os.makedirs(os.path.dirname(EVAL_DB_PATH),exist_ok=True)
conn=sqlite3.connect(EVAL_DB_PATH)
conn.execute("""
CREATE TABLE IF NOT EXISTS eval_results (
id INTEGER PRIMARY KEY AUTOINCREMENT,
run_timestamp TEXT NOT NULL,
finding_id TEXT NOT NULL,
relevancy_score REAL,
faithfulness_score REAL,
cwe_accuracy REAL,
overall_score REAL,
llm_model TEXT,
prompt_version TEXT,
generated_explanation TEXT,
reference_explanation TEXT,
passed INTEGER -- 1 if overall_score >= 0.6, else 0
)
""")
conn.commit()
conn.close()
def save_result(result:dict):
"""Save a single evaluation result to SQLite"""
conn=sqlite3.connect(EVAL_DB_PATH)
conn.execute("""
INSERT INTO eval_results(
run_timestamp, finding_id, relevancy_score, faithfulness_score, cwe_accuracy, overall_score, llm_model, prompt_version, generated_explanation, reference_explanation, passed
) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
""",(
result["run_timestamp"],
result["finding_id"],
result["relevancy_score"],
result["faithfulness_score"],
result["cwe_accuracy"],
result["overall_score"],
result.get("llm_model", "unknown"),
result.get("prompt_version", "unknown"),
result["generated_explanation"],
result["reference_explanation"],
1 if result["overall_score"] >= 0.6 else 0
))
conn.commit()
conn.close()
def get_eval_history(limit:int=50) -> list[dict]:
"""Fetch recent evaluation results"""
try:
conn=sqlite3.connect(EVAL_DB_PATH)
rows=conn.execute("""
SELECT run_timestamp,finding_id,relevancy_score,faithfulness_score,cwe_accuracy,overall_score,llm_model,prompt_version,passed
FROM eval_results
ORDER BY id DESC
LIMIT ?
""", (limit,)).fetchall()
conn.close()
return [
{
"run_timestamp": r[0],
"finding_id": r[1],
"relevancy_score": r[2],
"faithfulness_score": r[3],
"cwe_accuracy": r[4],
"overall_score": r[5],
"llm_model": r[6],
"prompt_version": r[7],
"passed": bool(r[8])
}
for r in rows
]
except Exception as e:
print(f"[eval] DB read error :{e}")
return []
def get_summary_stats()-> dict:
"""Compute aggregate stats across all eval runs"""
if not os.path.exists(EVAL_DB_PATH):
return []
try:
conn=sqlite3.connect(EVAL_DB_PATH)
row=conn.execute("""
SELECT
COUNT(*) AS total,
AVG(relevancy_score) AS avg_relevancy,
AVG(faithfulness_score) AS avg_faithfulness,
AVG(cwe_accuracy) AS avg_cwe,
AVG(overall_score) AS avg_overall,
SUM(passed) AS total_passed
FROM eval_results
""").fetchone()
conn.close()
if not row or row[0] ==0:
return {}
return{
"total_evaluations": row[0],
"avg_relevancy": round(row[1], 3),
"avg_faithfulness": round(row[2], 3),
"avg_cwe_accuracy": round(row[3], 3),
"avg_overall": round(row[4], 3),
"pass_rate": round(row[5] / row[0] * 100, 1)
}
except Exception as e:
print(f"[eval] Stats error: {e}")
return {}
#Neo4j helpers
def get_explained_findings_from_neo4j() ->list[dict]:
"""Fetch all findings with LLM explanations"""
driver=GraphDatabase.driver(NEO4J_URI,auth=(NEO4J_USER,NEO4J_PASSWORD))
try:
with driver.session() as session:
result=session.run("""
MATCH(n)
WHERE (n:Vulnerability OR n:Secret) AND n.explanation IS NOT NULL
RETURN CASE WHEN n:Vulnerability THEN n.id ELSE n.rule END AS finding_id,
n.explanation AS explanation,
n.why_dangerous AS why_dangerous,
n.fix AS fix,
n.cwe AS cwe,
n.llm_model AS llm_model,
n.llm_version AS prompt_version
""")
return [dict(r) for r in result]
finally:
driver.close()
#Core Eval function
def run_eval(verbose: bool = True) -> dict:
"""
Run eval against all findings which have LLM generated explanation in Neo4j and ground truth reference.
Returns summary statistics dict
"""
init_db()
gt_lookup={gt["finding_id"]: gt for gt in GROUND_TRUTH}
generated_findings=get_explained_findings_from_neo4j()
if not generated_findings:
print("[eval] No explained findings in Neo4j")
return {}
run_timestamp= datetime.now().isoformat()
results=[]
skipped=0
print(f"\n[eval] Evaluating {len(generated_findings)} findings against {len(gt_lookup)} ground truth entries...")
print("-" * 60)
for finding in generated_findings:
fid= finding.get("finding_id","")
gt=gt_lookup.get(fid)
if not gt:
skipped+=1
continue
generated_explanation= finding.get("explanation","")
generated_cwe=finding.get("cwe","")
if not generated_explanation:
skipped+=1
continue
relevancy=score_relevancy(
generated_explanation,gt["reference_explanation"]
)
faithfulness=score_faithfulness(
generated_explanation,gt["reference_fix"]
)
cwe_acc=score_cwe_accuracy(generated_cwe,gt["cwe"])
overall=compute_overall_score(relevancy,faithfulness, cwe_acc)
res={
"run_timestamp": run_timestamp,
"finding_id": fid,
"relevancy_score": relevancy,
"faithfulness_score": faithfulness,
"cwe_accuracy": cwe_acc,
"overall_score": overall,
"llm_model": finding.get("llm_model", "unknown"),
"prompt_version": finding.get("prompt_version", "unknown"),
"generated_explanation": generated_explanation,
"reference_explanation": gt["reference_explanation"]
}
save_result(res)
results.append(res)
if verbose:
status= "Pass" if overall >=0.6 else "FAIL"
print(f"{status} | {fid:<35} | relevancy={relevancy:.3f} | faithfulness={faithfulness:.3f} | cwe={cwe_acc:.1f} | overall={overall:.3f}")
print("-" * 60)
print(f"[eval] Evaluated: {len(results)} | Skipped (no ground truth): {skipped}")
#Summary stats
if results:
avg_relevancy = sum(r["relevancy_score"] for r in results) / len(results)
avg_faithfulness = sum(r["faithfulness_score"] for r in results) / len(results)
avg_cwe = sum(r["cwe_accuracy"] for r in results) / len(results)
avg_overall = sum(r["overall_score"] for r in results) / len(results)
pass_count = sum(1 for r in results if r["overall_score"] >= 0.6)
summary={
"total_evaluated": len(results),
"skipped": skipped,
"avg_relevancy": round(avg_relevancy, 3),
"avg_faithfulness": round(avg_faithfulness, 3),
"avg_cwe_accuracy": round(avg_cwe, 3),
"avg_overall": round(avg_overall, 3),
"pass_rate": round(pass_count / len(results) * 100, 1),
"passed": pass_count,
"failed": len(results) - pass_count
}
print(f"\nπ Summary:")
print(f" Avg Relevancy: {summary['avg_relevancy']:.3f}")
print(f" Avg Faithfulness: {summary['avg_faithfulness']:.3f}")
print(f" Avg CWE Accuracy: {summary['avg_cwe_accuracy']:.3f}")
print(f" Avg Overall: {summary['avg_overall']:.3f}")
print(f" Pass Rate: {summary['pass_rate']}% ({pass_count}/{len(results)})")
return summary
return {}
#CLI entry point
if __name__ == "__main__":
print("[eval] Starting VulnGraph LLM Evaluation Harness")
print("[eval] Metrics: Relevancy (50%) + Faithfulness (30%) + CWE Accuracy(20%)")
print("[eval] Pass threshold: Overall score >= 0.60\n")
summary= run_eval(verbose=True)
if summary:
print(f"\n[eval] Results saved to {EVAL_DB_PATH}") |