codejury 0.5.1__tar.gz → 0.6.0__tar.gz

{codejury-0.5.1 → codejury-0.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codejury
-Version: 0.5.1
+Version: 0.6.0
 Summary: General-purpose Application Security AI audit framework -- five-layer architecture, capabilities as first-class data
 Author: AISecLabs
 License-Expression: MIT
@@ -25,6 +25,7 @@ Provides-Extra: litellm
 Requires-Dist: litellm>=1.0; extra == "litellm"
 Provides-Extra: dev
 Requires-Dist: pytest>=8.0; extra == "dev"
+Requires-Dist: jsonschema>=4.0; extra == "dev"
 Dynamic: license-file
 
 # codejury
@@ -80,15 +81,32 @@ git diff | codejury audit --provider anthropic
 | `codejury audit [diff]` | Audit a unified diff from a file or stdin (`-`). |
 | `codejury scan <dir>` | Audit a whole directory tree, capability by capability. |
 | `codejury run <task>` | Run a named task preset (see [Tasks](#tasks)). |
-| `codejury eval` | Score the golden cases and report precision / recall. |
+| `codejury eval` | Score the golden cases; report precision / recall / F1, overall and per capability. |
 
 Shared flags: `--orchestrator {single,pipeline,debate,reflexion,challenge}`,
-`--provider {anthropic,openai,litellm}`, `--model`, `--format {text,markdown,json}`.
+`--provider {anthropic,openai,litellm}`, `--model`,
+`--format {text,markdown,json,sarif}`.
+
+`--format sarif` emits a SARIF 2.1.0 log (validates against the official schema)
+for CI and security dashboards: each problem with a code location becomes a
+result carrying its capability (as the rule id), CWE, and a precise location.
 
 Findings in known-noise categories (availability/DoS, rate limiting, memory safety
 outside C/C++) are dropped by versioned rules in
 `codejury/data/suppressions.yaml`; disable with `--no-suppress`.
 
+`codejury eval` takes `--dataset <dir>` (golden YAML directory), `--split <name>`
+(score only cases tagged with that `split:`, e.g. a held-out set), and
+`--format {text,json}` -- the JSON report is a stable schema (overall plus
+per-capability confusion matrix and precision / recall / F1).
+
+Runs are deterministic: providers query at temperature 0, and `audit` / `scan`
+cache each verdict on a hash of the normalized code, the in-scope capability
+versions, and the orchestration. Re-auditing unchanged code returns the recorded
+verdicts without re-querying the model; editing a capability YAML changes its
+fingerprint and invalidates affected entries. Pass `--no-cache` to always
+re-query.
+
 ```bash
 # Multi-round adversarial debate, rendered as Markdown
 git diff | codejury audit --orchestrator debate --format markdown - > report.md

{codejury-0.5.1 → codejury-0.6.0}/README.md

@@ -51,15 +51,32 @@ git diff | codejury audit --provider anthropic
 | `codejury audit [diff]` | Audit a unified diff from a file or stdin (`-`). |
 | `codejury scan <dir>` | Audit a whole directory tree, capability by capability. |
 | `codejury run <task>` | Run a named task preset (see [Tasks](#tasks)). |
-| `codejury eval` | Score the golden cases and report precision / recall. |
+| `codejury eval` | Score the golden cases; report precision / recall / F1, overall and per capability. |
 
 Shared flags: `--orchestrator {single,pipeline,debate,reflexion,challenge}`,
-`--provider {anthropic,openai,litellm}`, `--model`, `--format {text,markdown,json}`.
+`--provider {anthropic,openai,litellm}`, `--model`,
+`--format {text,markdown,json,sarif}`.
+
+`--format sarif` emits a SARIF 2.1.0 log (validates against the official schema)
+for CI and security dashboards: each problem with a code location becomes a
+result carrying its capability (as the rule id), CWE, and a precise location.
 
 Findings in known-noise categories (availability/DoS, rate limiting, memory safety
 outside C/C++) are dropped by versioned rules in
 `codejury/data/suppressions.yaml`; disable with `--no-suppress`.
 
+`codejury eval` takes `--dataset <dir>` (golden YAML directory), `--split <name>`
+(score only cases tagged with that `split:`, e.g. a held-out set), and
+`--format {text,json}` -- the JSON report is a stable schema (overall plus
+per-capability confusion matrix and precision / recall / F1).
+
+Runs are deterministic: providers query at temperature 0, and `audit` / `scan`
+cache each verdict on a hash of the normalized code, the in-scope capability
+versions, and the orchestration. Re-auditing unchanged code returns the recorded
+verdicts without re-querying the model; editing a capability YAML changes its
+fingerprint and invalidates affected entries. Pass `--no-cache` to always
+re-query.
+
 ```bash
 # Multi-round adversarial debate, rendered as Markdown
 git diff | codejury audit --orchestrator debate --format markdown - > report.md

{codejury-0.5.1 → codejury-0.6.0}/codejury/__init__.py

@@ -5,4 +5,9 @@ Domain knowledge lives in YAML capability files as a first-class citizen,
 aligned with OWASP ASVS.
 """
 
-__version__ = "0.0.0"
+from importlib.metadata import PackageNotFoundError, version
+
+try:
+    __version__ = version("codejury")
+except PackageNotFoundError:  # running from a source tree without an install
+    __version__ = "0.0.0"

{codejury-0.5.1 → codejury-0.6.0}/codejury/agents/verifier.py

@@ -93,15 +93,27 @@ def _build_prompt(path: str, content: str, cap: Capability, context: str = "") -
     )
 
 
+def _anti_pattern_cwes(cap: Capability) -> dict[str, str]:
+    """Map anti_pattern id -> CWE, so a verdict can inherit the CWE it matched."""
+    return {
+        p.id: p.cwe
+        for sub in cap.sub_capabilities.values()
+        for p in sub.anti_patterns
+        if p.cwe
+    }
+
+
 def _parse_verdicts(text: str, cap: Capability) -> list[Verdict]:
     obj = extract_json_object(text)
     if not obj:
         return []
+    cwe_by_id = _anti_pattern_cwes(cap)
     out: list[Verdict] = []
     for v in obj.get("verdicts", []):
         if not isinstance(v, dict):
             continue
         sub = str(v.get("sub_capability", "")).strip()
+        matched_anti = str_list(v.get("matched_anti"))
         out.append(
             Verdict(
                 capability=f"{cap.id}.{sub}" if sub else cap.id,
@@ -109,7 +121,8 @@ def _parse_verdicts(text: str, cap: Capability) -> list[Verdict]:
                 status=one_of(v.get("status"), _VALID_STATUS, "UNKNOWN"),
                 reasoning=str(v.get("reasoning", "")),
                 matched_correct=str_list(v.get("matched_correct")),
-                matched_anti=str_list(v.get("matched_anti")),
+                matched_anti=matched_anti,
+                cwe=next((cwe_by_id[a] for a in matched_anti if a in cwe_by_id), ""),
                 evidence=to_evidence(v.get("evidence")),
                 confidence=to_float(v.get("confidence"), 0.5),
             )

{codejury-0.5.1 → codejury-0.6.0}/codejury/assembly.py

@@ -16,6 +16,7 @@ from codejury.domain.artifact import CodeArtifact
 from codejury.domain.capability import Capability
 from codejury.domain.context import AnalysisContext
 from codejury.domain.result import AnalysisResult
+from codejury.infrastructure.cache import VerdictCache, verdict_key
 from codejury.orchestrators.base import Orchestrator
 from codejury.orchestrators.challenge import ChallengeOrchestrator
 from codejury.orchestrators.debate import DebateOrchestrator
@@ -75,17 +76,38 @@ def build_orchestration(
     return verifier, SingleOrchestrator()
 
 
+def orchestration_descriptor(strategy: str, model: str, max_tokens: int) -> str:
+    """The non-code, non-capability inputs that affect a verdict, as a cache tag."""
+    return f"{strategy}|{model}|{max_tokens}"
+
+
 def run_over_artifacts(
     artifacts: list[CodeArtifact],
     capabilities: list[Capability],
     agents: dict[str, Agent],
     orchestrator: Orchestrator,
+    *,
+    cache: VerdictCache | None = None,
+    orchestration: str = "",
 ) -> list[tuple[str, AnalysisResult]]:
-    """Run the orchestration over each artifact, returning (path, result) per artifact."""
+    """Run the orchestration over each artifact, returning (path, result) per artifact.
+
+    When ``cache`` is given, an unchanged artifact returns its recorded result
+    instead of re-running the orchestrator (determinism, invariant 2).
+    """
     results = []
     for artifact in artifacts:
+        if cache is not None:
+            key = verdict_key(artifact, capabilities, orchestration=orchestration)
+            hit = cache.get(key)
+            if hit is not None:
+                results.append((artifact.path, hit))
+                continue
         ctx = AnalysisContext(artifact=artifact, capabilities=capabilities)
-        results.append((artifact.path, orchestrator.run(agents, ctx)))
+        result = orchestrator.run(agents, ctx)
+        if cache is not None:
+            cache.put(key, result)
+        results.append((artifact.path, result))
     return results
 
 
@@ -94,5 +116,11 @@ def run_over_source(
     capabilities: list[Capability],
     agents: dict[str, Agent],
     orchestrator: Orchestrator,
+    *,
+    cache: VerdictCache | None = None,
+    orchestration: str = "",
 ) -> list[tuple[str, AnalysisResult]]:
-    return run_over_artifacts(source.list_artifacts(), capabilities, agents, orchestrator)
+    return run_over_artifacts(
+        source.list_artifacts(), capabilities, agents, orchestrator,
+        cache=cache, orchestration=orchestration,
+    )

{codejury-0.5.1 → codejury-0.6.0}/codejury/cli.py

@@ -9,6 +9,7 @@ library, backed by the Anthropic provider, under a chosen orchestration strategy
 from __future__ import annotations
 
 import argparse
+import json
 import os
 import sys
 
@@ -21,6 +22,7 @@ from codejury.assembly import (
     STRATEGIES,
     build_orchestration,
     make_provider,
+    orchestration_descriptor,
     run_over_artifacts,
     run_over_source,
 )
@@ -29,11 +31,12 @@ from codejury.domain.capability import Capability, load_capabilities
 from codejury.domain.context import AnalysisContext
 from codejury.domain.observation import Observation
 from codejury.domain.result import AnalysisResult
-from codejury.evaluation import Metrics, evaluate, load_cases
+from codejury.evaluation import EvalReport, evaluate, load_cases
+from codejury.infrastructure.cache import VerdictCache
 from codejury.orchestrators.single import SingleOrchestrator
 from codejury.providers.base import Provider
 from codejury.providers.mock import MockProvider
-from codejury.reporting import to_json, to_markdown
+from codejury.reporting import to_json, to_markdown, to_sarif
 from codejury.resources import CAPABILITIES_DIR, GOLDEN_DIR, SUPPRESSIONS_FILE, TASKS_DIR
 from codejury.suppression import filter_results, load_suppressions
 from codejury.integrations.github import build_review, parse_pr_ref, post_review
@@ -43,7 +46,7 @@ from codejury.sources.repo import RepoSource
 from codejury.tasks.base import run_task
 from codejury.tasks.registry import load_tasks
 
-_FORMATS = ("text", "markdown", "json")
+_FORMATS = ("text", "markdown", "json", "sarif")
 
 
 def dry_run() -> AnalysisResult:
@@ -69,10 +72,14 @@ def audit(
     model: str,
     max_tokens: int = 2048,
     strategy: str = "single",
+    cache: VerdictCache | None = None,
 ) -> list[tuple[str, AnalysisResult]]:
     """Audit each changed file in `diff_text`, returning (path, result) per file."""
     agents, orchestrator = build_orchestration(strategy, provider=provider, model=model, max_tokens=max_tokens)
-    return run_over_source(DiffSource(diff_text), capabilities, agents, orchestrator)
+    return run_over_source(
+        DiffSource(diff_text), capabilities, agents, orchestrator,
+        cache=cache, orchestration=orchestration_descriptor(strategy, model, max_tokens),
+    )
 
 
 def scan(
@@ -87,6 +94,7 @@ def scan(
     max_chars: int = 200_000,
     with_callers: bool = False,
     with_callees: bool = False,
+    cache: VerdictCache | None = None,
 ) -> list[tuple[str, AnalysisResult]]:
     """Audit every matching file in a directory tree, returning (path, result) per artifact."""
     source = RepoSource(
@@ -103,7 +111,10 @@ def scan(
         file=sys.stderr,
     )
     agents, orchestrator = build_orchestration(strategy, provider=provider, model=model, max_tokens=max_tokens)
-    return run_over_artifacts(artifacts, capabilities, agents, orchestrator)
+    return run_over_artifacts(
+        artifacts, capabilities, agents, orchestrator,
+        cache=cache, orchestration=orchestration_descriptor(strategy, model, max_tokens),
+    )
 
 
 def _render_dry_run(result: AnalysisResult) -> str:
@@ -142,7 +153,7 @@ def _render_observation(o: Observation) -> str:
 
 
 def _render_results(fmt: str, results: list[tuple[str, AnalysisResult]]) -> str:
-    return {"text": _render_audit, "markdown": to_markdown, "json": to_json}[fmt](results)
+    return {"text": _render_audit, "markdown": to_markdown, "json": to_json, "sarif": to_sarif}[fmt](results)
 
 
 def _maybe_suppress(results: list[tuple[str, AnalysisResult]], enabled: bool) -> list[tuple[str, AnalysisResult]]:
@@ -189,11 +200,16 @@ def _maybe_post_github(ref: str | None, results: list[tuple[str, AnalysisResult]
         print(f"github review failed: {exc}", file=sys.stderr)
 
 
-def _render_metrics(m: Metrics) -> str:
-    return (
-        f"cases: {m.total}  (tp={m.tp} fp={m.fp} tn={m.tn} fn={m.fn})\n"
-        f"precision: {m.precision:.2f}  recall: {m.recall:.2f}  accuracy: {m.accuracy:.2f}"
-    )
+def _render_eval(report: EvalReport) -> str:
+    def line(label: str, m) -> str:
+        return (
+            f"{label:<20} tp={m.tp} fp={m.fp} tn={m.tn} fn={m.fn}  "
+            f"P={m.precision:.2f} R={m.recall:.2f} F1={m.f1:.2f}"
+        )
+
+    lines = [line(f"overall ({report.overall.total} cases)", report.overall)]
+    lines += [line(cap, m) for cap, m in sorted(report.by_capability.items())]
+    return "\n".join(lines)
 
 
 def _read_diff(path: str) -> str:
@@ -221,6 +237,7 @@ def main(argv: list[str] | None = None) -> int:
     audit_p.add_argument("--api-base", default=DEFAULT_API_BASE, help="provider base URL (env: CODEJURY_API_BASE)")
     audit_p.add_argument("--api-key", default=DEFAULT_API_KEY, help="provider API key (env: CODEJURY_API_KEY)")
     audit_p.add_argument("--no-suppress", action="store_true", help="disable the known-noise suppression filter")
+    audit_p.add_argument("--no-cache", action="store_true", help="bypass the verdict cache (always re-query the model)")
     audit_p.add_argument("--fail-on", choices=_FAIL_ON, default=None, dest="fail_on", help="exit 1 if a finding at/above this severity is found")
     audit_p.add_argument("--github", default=None, help="post a PR review: owner/repo#number (needs GITHUB_TOKEN)")
 
@@ -244,6 +261,7 @@ def main(argv: list[str] | None = None) -> int:
     scan_p.add_argument("--api-base", default=DEFAULT_API_BASE, help="provider base URL (env: CODEJURY_API_BASE)")
     scan_p.add_argument("--api-key", default=DEFAULT_API_KEY, help="provider API key (env: CODEJURY_API_KEY)")
     scan_p.add_argument("--no-suppress", action="store_true", help="disable the known-noise suppression filter")
+    scan_p.add_argument("--no-cache", action="store_true", help="bypass the verdict cache (always re-query the model)")
     scan_p.add_argument("--fail-on", choices=_FAIL_ON, default=None, dest="fail_on", help="exit 1 if a finding at/above this severity is found")
 
     run_p = sub.add_parser("run", help="run a named task preset against a unified diff")
@@ -256,9 +274,11 @@ def main(argv: list[str] | None = None) -> int:
     run_p.add_argument("--fail-on", choices=_FAIL_ON, default=None, dest="fail_on", help="exit 1 if a finding at/above this severity is found")
 
     eval_p = sub.add_parser("eval", help="score golden cases and report precision/recall")
-    eval_p.add_argument("--golden", default=GOLDEN_DIR, help="golden case YAML directory")
+    eval_p.add_argument("--dataset", default=GOLDEN_DIR, help="golden case YAML directory")
+    eval_p.add_argument("--split", default=None, help="only score cases whose 'split' matches (e.g. held-out)")
     eval_p.add_argument("--capabilities", default=CAPABILITIES_DIR, help="capability YAML directory")
     eval_p.add_argument("--provider", choices=PROVIDERS, default="anthropic")
+    eval_p.add_argument("--format", choices=("text", "json"), default="text", dest="fmt")
     eval_p.add_argument("--model", default=DEFAULT_MODEL)
     eval_p.add_argument("--api-base", default=DEFAULT_API_BASE, help="provider base URL (env: CODEJURY_API_BASE)")
     eval_p.add_argument("--api-key", default=DEFAULT_API_KEY, help="provider API key (env: CODEJURY_API_KEY)")
@@ -275,6 +295,7 @@ def main(argv: list[str] | None = None) -> int:
             model=args.model,
             max_tokens=args.max_tokens,
             strategy=args.orchestrator,
+            cache=None if args.no_cache else VerdictCache(),
         )
         results = _maybe_suppress(results, not args.no_suppress)
         print(_render_results(args.fmt, results))
@@ -298,6 +319,7 @@ def main(argv: list[str] | None = None) -> int:
             max_chars=args.max_chars,
             with_callers=args.callers,
             with_callees=args.callees,
+            cache=None if args.no_cache else VerdictCache(),
         )
         results = _maybe_suppress(results, not args.no_suppress)
         print(_render_results(args.fmt, results))
@@ -317,8 +339,8 @@ def main(argv: list[str] | None = None) -> int:
 
     if args.command == "eval":
         try:
-            metrics = evaluate(
-                load_cases(args.golden),
+            report = evaluate(
+                load_cases(args.dataset, split=args.split),
                 load_capabilities(args.capabilities),
                 provider=make_provider(args.provider, api_key=args.api_key, api_base=args.api_base),
                 model=args.model,
@@ -328,7 +350,7 @@ def main(argv: list[str] | None = None) -> int:
             # as one line, not a traceback (audit gets this via the orchestrator).
             print(f"eval failed: {exc}")
             return 1
-        print(_render_metrics(metrics))
+        print(json.dumps(report.to_dict(), indent=2) if args.fmt == "json" else _render_eval(report))
         return 0
 
     if args.command in (None, "dry-run"):

{codejury-0.5.1 → codejury-0.6.0}/codejury/data/capabilities/dependency_config.yaml

@@ -46,7 +46,34 @@ sub_capabilities:
         signals: ["admin:admin", "password=admin", "changeme"]
         why_bad: Default credentials are public knowledge and trivially abused
 
+  transport_security:
+    correct_patterns:
+      - id: TLS-OK-1
+        description: >-
+          Leave TLS certificate verification at its secure default -- verify omitted or
+          verify=True, the default SSL context, hostname checking on
+        signals: ["verify=True", "create_default_context", "requests.get(", "requests.post("]
+        why_ok: >-
+          The secure default validates the certificate chain and hostname. An https:// call
+          that does not disable verification is fine; do not flag it just for making a
+          request or for omitting verify.
+
+    anti_patterns:
+      - id: TLS-BAD-1
+        cwe: CWE-295
+        severity: HIGH
+        description: >-
+          Disable TLS certificate or hostname verification -- verify=False, CERT_NONE,
+          check_hostname=False, or an unverified SSL context
+        signals: ["verify=False", "CERT_NONE", "check_hostname = False", "_create_unverified_context"]
+        why_bad: An unverified TLS connection is open to a man-in-the-middle despite https://
+        example_bad: |
+          requests.get("https://api.partner.com/data", verify=False)
+        example_good: |
+          requests.get("https://api.partner.com/data")  # verify defaults to True
+
 trigger_signals:
   - dependency manifests and lock files
   - install or bootstrap scripts fetching remote code
   - file permission, bucket ACL, or default credential settings
+  - TLS client calls that set verify or build a custom SSL context

{codejury-0.5.1 → codejury-0.6.0}/codejury/data/capabilities/input_validation.yaml

@@ -105,7 +105,67 @@ sub_capabilities:
           if not target.is_relative_to(UPLOAD_DIR):
               raise ValueError("path escapes upload dir")
 
+  ssrf:
+    correct_patterns:
+      - id: SSRF-OK-1
+        description: Validate the request URL's host against an allowlist before fetching it
+        signals: ["urlparse(", ".hostname", "ALLOWED", "allowlist"]
+        why_ok: An attacker cannot redirect the fetch to an internal target the list omits
+
+      - id: SSRF-OK-2
+        description: >-
+          Fetch a URL that is not attacker-controlled -- a constant, a value from trusted
+          config, or an operator-supplied argument
+        why_ok: >-
+          SSRF needs an external attacker to control the destination. A constant URL or one
+          from trusted config is not a finding, even though it goes through a fetch call.
+
+    anti_patterns:
+      - id: SSRF-BAD-1
+        cwe: CWE-918
+        severity: HIGH
+        description: >-
+          Fetch a URL taken from externally controlled input (HTTP request, form, query, or
+          message field) without validating its host against an allowlist. NOT this: a
+          constant URL, one from trusted config, or an operator-supplied argument.
+        signals: ["requests.get(", "urllib.request.urlopen(", "httpx.", "request.args", "request.json"]
+        why_bad: >-
+          The server makes the request, so attacker input reaches internal-only targets --
+          cloud metadata, localhost admin ports, internal APIs behind the firewall.
+        example_bad: |
+          requests.get(request.args["url"]).text
+        example_good: |
+          if urlparse(url).hostname not in ALLOWED_HOSTS:
+              raise ValueError("host not allowed")
+          requests.get(url).text
+
+  insecure_deserialization:
+    correct_patterns:
+      - id: DESER-OK-1
+        description: >-
+          Parse untrusted input with a data-only parser -- json.loads or yaml.safe_load --
+          that cannot instantiate arbitrary objects
+        signals: ["json.loads", "yaml.safe_load"]
+        why_ok: A data-only parser builds plain structures and has no code-execution path
+
+    anti_patterns:
+      - id: DESER-BAD-1
+        cwe: CWE-502
+        severity: CRITICAL
+        description: >-
+          Deserialize externally controlled bytes with an object-constructing deserializer --
+          pickle, marshal, yaml.load (unsafe Loader), or jsonpickle. NOT this: a data-only
+          parser like json.loads or yaml.safe_load.
+        signals: ["pickle.loads", "pickle.load(", "yaml.load(", "marshal.loads", "jsonpickle.decode"]
+        why_bad: These reconstruct arbitrary objects, so crafted input runs code on unpickle
+        example_bad: |
+          pickle.loads(base64.b64decode(request.data))
+        example_good: |
+          json.loads(request.data)
+
 trigger_signals:
   - raw SQL strings or cursor.execute calls appear
   - imports of os, subprocess, or shlex with process execution
   - file paths built from request, form, or query parameters
+  - outbound HTTP fetches (requests, urllib, httpx) to a non-constant URL
+  - deserialization calls (pickle, yaml.load, marshal) on external input

codejury-0.6.0/codejury/data/golden/authn_weak_hash_indirect_vuln.yaml

@@ -0,0 +1,14 @@
+# Adversarial positive: weak password hash hidden behind hashlib.new(variable).
+capability: authn
+vulnerable: true
+expected_verdict: VULNERABLE
+cwe: CWE-916
+source: synthetic
+notes: >
+  hashlib.new("md5") is the same weak, unsalted, fast hash as hashlib.md5(), just
+  reached through a variable algorithm name. Unsuitable for password storage
+  (needs bcrypt/scrypt/argon2). The indirection should not hide it.
+code: |
+  def hash_pw(pw):
+      algo = "md5"
+      return hashlib.new(algo, pw.encode()).hexdigest()

codejury-0.6.0/codejury/data/golden/business_logic_price_tamper_vuln.yaml

@@ -0,0 +1,14 @@
+capability: business_logic
+vulnerable: true
+expected_verdict: VULNERABLE
+cwe: CWE-602
+source: synthetic
+split: held-out
+notes: >
+  Quantity and unit price come straight from the request and are never checked
+  server-side. A negative quantity credits the customer; a client-set price lets
+  them pay anything. Price must come from the catalog and quantity must be > 0.
+code: |
+  def add_to_order(order, item_id, quantity, unit_price):
+      order.lines.append((item_id, quantity, unit_price))
+      order.total += quantity * unit_price

codejury-0.6.0/codejury/data/golden/business_logic_server_checked_safe.yaml

@@ -0,0 +1,15 @@
+capability: business_logic
+vulnerable: false
+expected_verdict: SECURE
+cwe: ""
+source: synthetic
+notes: >
+  Quantity is validated as positive and the price is looked up server-side from
+  the catalog, so the client cannot tamper with either.
+code: |
+  def add_to_order(order, item_id, quantity):
+      if quantity < 1:
+          raise ValueError("quantity must be positive")
+      unit_price = catalog.price_of(item_id)   # trusted server-side price
+      order.lines.append((item_id, quantity, unit_price))
+      order.total += quantity * unit_price

codejury-0.6.0/codejury/data/golden/cmdi_fixed_argv_safe.yaml

@@ -0,0 +1,22 @@
+# False-positive-prone negative: a subprocess call driven by a request value
+# looks like command injection, but the value only selects a fixed argv from a
+# table and never reaches a shell.
+capability: input_validation
+vulnerable: false
+expected_verdict: SECURE
+cwe: ""
+source: synthetic
+split: held-out
+notes: >
+  `name` only indexes a fixed dict of argument lists (KeyError on anything
+  unknown); no user string is interpolated into a command and shell=False.
+  Flagging this as command injection is a false positive.
+code: |
+  ACTIONS = {
+      "start": ["systemctl", "start", "web"],
+      "stop": ["systemctl", "stop", "web"],
+  }
+
+  def run_action(name):
+      argv = ACTIONS[name]
+      subprocess.run(argv, shell=False)

codejury-0.6.0/codejury/data/golden/data_protection_plaintext_pii_vuln.yaml

@@ -0,0 +1,14 @@
+capability: data_protection
+vulnerable: true
+expected_verdict: VULNERABLE
+cwe: CWE-312
+source: synthetic
+notes: >
+  The full card number and the CVV are written to the database in cleartext.
+  The CVV must never be stored at all, and the PAN must be tokenized/encrypted.
+code: |
+  def save_card(user_id, card_number, cvv):
+      db.execute(
+          "INSERT INTO cards (user_id, pan, cvv) VALUES (?, ?, ?)",
+          (user_id, card_number, cvv),
+      )

codejury-0.6.0/codejury/data/golden/data_protection_tokenized_safe.yaml

@@ -0,0 +1,16 @@
+capability: data_protection
+vulnerable: false
+expected_verdict: SECURE
+cwe: ""
+source: synthetic
+split: held-out
+notes: >
+  The PAN is replaced by a vault token before storage; the CVV is never
+  persisted. Nothing sensitive reaches the database.
+code: |
+  def save_card(user_id, card_number):
+      token = vault.tokenize(card_number)   # store a token, never the PAN
+      db.execute(
+          "INSERT INTO cards (user_id, token) VALUES (?, ?)",
+          (user_id, token),
+      )

codejury-0.6.0/codejury/data/golden/dependency_config_tls_verify_off_vuln.yaml

@@ -0,0 +1,11 @@
+capability: dependency_config
+vulnerable: true
+expected_verdict: VULNERABLE
+cwe: CWE-295
+source: synthetic
+notes: >
+  TLS certificate verification is disabled, so the connection is open to a
+  man-in-the-middle despite the https:// scheme.
+code: |
+  def fetch_partner_data():
+      return requests.get("https://api.partner.com/data", verify=False).json()

codejury-0.6.0/codejury/data/golden/dependency_config_tls_verify_on_safe.yaml

@@ -0,0 +1,11 @@
+capability: dependency_config
+vulnerable: false
+expected_verdict: SECURE
+cwe: ""
+source: synthetic
+notes: >
+  Certificate verification is left at its secure default (True) and a timeout is
+  set. Nothing weakens the TLS configuration.
+code: |
+  def fetch_partner_data():
+      return requests.get("https://api.partner.com/data", timeout=10).json()

codejury-0.6.0/codejury/data/golden/deserialize_json_safe.yaml

@@ -0,0 +1,13 @@
+# False-positive-prone negative: deserializing request data looks dangerous,
+# but json.loads only builds data -- it cannot execute code like pickle.
+capability: input_validation
+vulnerable: false
+expected_verdict: SECURE
+cwe: ""
+source: synthetic
+notes: >
+  json.loads parses untrusted input into plain data structures with no code
+  execution path. Treating it like insecure deserialization is a false positive.
+code: |
+  def load_state(request):
+      return json.loads(request.data)

codejury-0.6.0/codejury/data/golden/deserialize_pickle_vuln.yaml

@@ -0,0 +1,12 @@
+capability: input_validation
+vulnerable: true
+expected_verdict: VULNERABLE
+cwe: CWE-502
+source: synthetic
+notes: >
+  pickle.loads on attacker-supplied bytes executes arbitrary code during
+  unpickling -- remote code execution.
+code: |
+  def load_state(request):
+      blob = base64.b64decode(request.data)
+      return pickle.loads(blob)

codejury-0.6.0/codejury/data/golden/error_logging_redacted_safe.yaml

@@ -0,0 +1,11 @@
+capability: error_logging
+vulnerable: false
+expected_verdict: SECURE
+cwe: ""
+source: synthetic
+notes: >
+  Only the non-sensitive username is logged; the password never enters the log.
+code: |
+  def authenticate(username, password):
+      logger.info("login attempt user=%s", username)
+      return backend.check(username, password)

codejury-0.6.0/codejury/data/golden/error_logging_secret_leak_vuln.yaml

@@ -0,0 +1,12 @@
+capability: error_logging
+vulnerable: true
+expected_verdict: VULNERABLE
+cwe: CWE-532
+source: synthetic
+notes: >
+  The cleartext password is written to the log. Anyone with log access (or a log
+  shipper, or an aggregation backend) now has the credential.
+code: |
+  def authenticate(username, password):
+      logger.info("login attempt user=%s password=%s", username, password)
+      return backend.check(username, password)