attackmap 0.1.0__py3-none-any.whl

attackmap/defensive_review.py

@@ -0,0 +1,597 @@
+from __future__ import annotations
+
+import re
+
+from .models import (
+    Asset,
+    AttackPath,
+    AttackSurface,
+    Control,
+    DetectionOpportunity,
+    Finding,
+    Insight,
+    ScanResult,
+)
+from .security_overlay import SecurityOverlay, build_security_overlay
+
+LOW_QUALITY_SEGMENTS = ("/tests/", "/__tests__/", "/fixtures/", "/mocks/", "/examples/")
+
+
+def _severity_rank(value: str) -> int:
+    return {"high": 0, "medium": 1, "low": 2}.get(value, 3)
+
+
+def _surface_risk_rank(value: str) -> int:
+    return {"high": 0, "medium": 1, "low": 2}.get(value, 3)
+
+
+def _confidence_rank(value: str) -> float:
+    return {"high": 3.0, "medium": 2.0, "low": 1.0}.get(value, 1.0)
+
+
+def _extract_numeric_confidence(text: str) -> float | None:
+    match = re.search(r"confidence=([0-9]+(?:\.[0-9]+)?)", text.lower())
+    if match is None:
+        return None
+    try:
+        return float(match.group(1))
+    except ValueError:
+        return None
+
+
+def _contains_any(text: str, keywords: tuple[str, ...]) -> bool:
+    lowered = text.lower()
+    return any(keyword in lowered for keyword in keywords)
+
+
+def _is_low_quality_source(path_or_text: str) -> bool:
+    normalized = path_or_text.replace("\\", "/").lower()
+    return any(segment in f"/{normalized}/" for segment in LOW_QUALITY_SEGMENTS)
+
+
+def _is_protocol_derived_surface(surface: AttackSurface) -> bool:
+    file_lower = surface.file.lower()
+    route_lower = surface.route.lower()
+    return "lexicon" in file_lower or "/xrpc/" in route_lower or "atproto_" in " ".join(surface.auth_signals).lower()
+
+
+def _entrypoint_counts(attack_surfaces: list[AttackSurface]) -> tuple[int, int, int, int]:
+    observed_runtime_public = 0
+    protocol_derived = 0
+    internal_only = 0
+    low_quality = 0
+    for surface in attack_surfaces:
+        if _is_low_quality_source(surface.file):
+            low_quality += 1
+            continue
+        if surface.exposure == "internal":
+            internal_only += 1
+            continue
+        if _is_protocol_derived_surface(surface):
+            protocol_derived += 1
+            continue
+        if surface.exposure == "public":
+            observed_runtime_public += 1
+    return observed_runtime_public, protocol_derived, internal_only, low_quality
+
+
+def _surface_provenance(surface: AttackSurface) -> str:
+    if _is_low_quality_source(surface.file):
+        return "low_quality"
+    if _is_protocol_derived_surface(surface):
+        return "protocol_derived"
+    if surface.exposure == "public":
+        return "observed_runtime"
+    return "other"
+
+
+def _provenance_breakdown(surfaces: list[AttackSurface]) -> str:
+    if not surfaces:
+        return "observed_runtime=0%, protocol_derived=0%, low_quality=0%"
+    counts = {"observed_runtime": 0, "protocol_derived": 0, "low_quality": 0}
+    for surface in surfaces:
+        key = _surface_provenance(surface)
+        if key in counts:
+            counts[key] += 1
+    total = max(len(surfaces), 1)
+    observed_pct = round((counts["observed_runtime"] / total) * 100)
+    protocol_pct = round((counts["protocol_derived"] / total) * 100)
+    low_quality_pct = round((counts["low_quality"] / total) * 100)
+    return (
+        f"observed_runtime={observed_pct}%, "
+        f"protocol_derived={protocol_pct}%, "
+        f"low_quality={low_quality_pct}%"
+    )
+
+
+def _provenance_counts(surfaces: list[AttackSurface]) -> dict[str, int]:
+    counts = {"observed_runtime": 0, "protocol_derived": 0, "low_quality": 0}
+    for surface in surfaces:
+        key = _surface_provenance(surface)
+        if key in counts:
+            counts[key] += 1
+    return counts
+
+
+def _recommendation_basis_label(surfaces: list[AttackSurface]) -> str:
+    if not surfaces:
+        return "inferred-weak"
+    counts = _provenance_counts(surfaces)
+    total = max(sum(counts.values()), 1)
+    observed_pct = (counts["observed_runtime"] / total) * 100
+    protocol_pct = (counts["protocol_derived"] / total) * 100
+    low_quality_pct = (counts["low_quality"] / total) * 100
+    if observed_pct >= 50:
+        return "observed"
+    if protocol_pct >= 50:
+        return "inferred-protocol"
+    if low_quality_pct >= 50:
+        return "low-quality-evidence"
+    return "mixed-evidence"
+
+
+def _top_score_reasons(factors: dict[str, float], top_n: int = 3) -> str:
+    ranked = sorted(factors.items(), key=lambda item: item[1], reverse=True)
+    top = [f"{name}={value:.1f}" for name, value in ranked[:top_n] if value > 0]
+    return ", ".join(top)
+
+
+def _surface_priority(surface: AttackSurface) -> tuple[float, dict[str, float]]:
+    exposure_score = {"public": 3.0, "unknown": 2.0, "internal": 1.0}.get(surface.exposure, 1.0)
+    privilege_score = {
+        "admin": 3.0,
+        "auth": 3.0,
+        "webhook": 2.0,
+        "upload": 2.0,
+        "public_api": 1.5,
+        "internal": 1.0,
+        "health": 0.5,
+    }.get(surface.category, 1.0)
+    reachability_score = 2.0 if surface.exposure == "public" else 1.0
+    trust_boundary_score = (2.0 if surface.data_store_interaction else 0.0) + (2.0 if surface.outbound_integration else 0.0)
+    chain_depth_score = 1.0 + (1.0 if surface.data_store_interaction and surface.outbound_integration else 0.0)
+    confidence_score = 2.0 if surface.auth_signals else 1.0
+    weighted = {
+        "exposure": 2.0 * exposure_score,
+        "privilege": 2.0 * privilege_score,
+        "reachability": 1.5 * reachability_score,
+        "trust_boundary": 1.5 * trust_boundary_score,
+        "chain_depth": 1.0 * chain_depth_score,
+        "confidence": 1.0 * confidence_score,
+    }
+    if _is_low_quality_source(surface.file):
+        weighted["source_quality_penalty"] = -8.0
+    elif _is_protocol_derived_surface(surface):
+        weighted["source_quality_penalty"] = -3.0
+    elif surface.exposure == "public":
+        weighted["source_quality_bonus"] = 1.0
+    return round(sum(weighted.values()), 2), weighted
+
+
+def _path_priority(path: AttackPath) -> tuple[float, dict[str, float]]:
+    text = f"{path.name} {path.impact} {' '.join(path.steps)}"
+    exposure_score = 3.0 if _contains_any(text, ("attacker reaches", "public", "/xrpc/")) else 2.0
+    privilege_score = 1.0 + (
+        2.0
+        if _contains_any(text, ("admin", "privileged", "authorization", "authz", "token", "identity"))
+        else 0.5
+    )
+    reachability_score = 2.0 if any(step.lower().startswith("entry:") for step in path.steps) else 1.0
+    trust_boundary_score = 1.0 + sum(
+        1.0
+        for step in path.steps
+        if _contains_any(step, ("propagation", "service", "edge", "external", "config risk", "sink"))
+    )
+    chain_depth_score = max(1.0, min(len(path.steps), 8) / 2.0)
+    evidence_text = " ".join(path.steps)
+    numeric_confidence = _extract_numeric_confidence(evidence_text)
+    confidence_score = 1.5 + (numeric_confidence if numeric_confidence is not None else 0.5)
+    weighted = {
+        "exposure": 2.0 * exposure_score,
+        "privilege": 2.0 * privilege_score,
+        "reachability": 1.5 * reachability_score,
+        "trust_boundary": 2.0 * trust_boundary_score,
+        "chain_depth": 1.2 * chain_depth_score,
+        "confidence": 1.5 * confidence_score,
+    }
+    if _is_low_quality_source(text):
+        weighted["source_quality_penalty"] = -10.0
+    return round(sum(weighted.values()), 2), weighted
+
+
+def _related_surfaces_for_finding(finding: Finding, attack_surfaces: list[AttackSurface]) -> list[AttackSurface]:
+    combined = f"{finding.title} {' '.join(finding.evidence)}".lower()
+    related = [
+        surface
+        for surface in attack_surfaces
+        if surface.route.lower() in combined
+        or surface.file.lower() in combined
+        or surface.category.lower() in combined
+    ]
+    return related
+
+
+def _related_paths_for_finding(finding: Finding, attack_paths: list[AttackPath], surfaces: list[AttackSurface]) -> list[AttackPath]:
+    if not attack_paths:
+        return []
+    combined = f"{finding.title} {' '.join(finding.evidence)}".lower()
+    surface_tokens = [surface.route.lower() for surface in surfaces if surface.route]
+    related = [
+        path
+        for path in attack_paths
+        if path.name.lower() in combined
+        or any(step.lower() in combined for step in path.steps[:2])
+        or any(token in " ".join(path.steps).lower() for token in surface_tokens)
+    ]
+    return related
+
+
+def _finding_priority(finding: Finding, attack_surfaces: list[AttackSurface], attack_paths: list[AttackPath]) -> tuple[float, dict[str, float]]:
+    surfaces = _related_surfaces_for_finding(finding, attack_surfaces)
+    paths = _related_paths_for_finding(finding, attack_paths, surfaces)
+
+    exposure_score = max((3.0 if surface.exposure == "public" else 1.0 for surface in surfaces), default=1.0)
+    privilege_score = max(
+        (
+            3.0
+            if surface.category in {"admin", "auth"}
+            else 2.0
+            if surface.category in {"webhook", "upload"}
+            else 1.0
+            for surface in surfaces
+        ),
+        default=1.0,
+    )
+    if _contains_any(finding.title, ("privilege", "admin", "auth", "token", "identity")):
+        privilege_score = max(privilege_score, 2.5)
+    reachability_score = max((2.0 if surface.exposure == "public" else 1.0 for surface in surfaces), default=1.0)
+    trust_boundary_score = max(
+        (
+            (2.0 if surface.data_store_interaction else 0.0) + (2.0 if surface.outbound_integration else 0.0)
+            for surface in surfaces
+        ),
+        default=1.0,
+    )
+    chain_depth_score = max((min(len(path.steps), 8) / 2.0 for path in paths), default=0.5)
+    numeric_confidence = max(
+        (
+            value
+            for value in (_extract_numeric_confidence(item) for item in finding.evidence)
+            if value is not None
+        ),
+        default=None,
+    )
+    confidence_score = _confidence_rank(finding.confidence) + (numeric_confidence if numeric_confidence is not None else 0.0)
+
+    weighted = {
+        "exposure": 2.0 * exposure_score,
+        "privilege": 2.0 * privilege_score,
+        "reachability": 1.5 * reachability_score,
+        "trust_boundary": 2.0 * trust_boundary_score,
+        "chain_depth": 1.2 * chain_depth_score,
+        "confidence": 1.5 * confidence_score,
+    }
+    if not surfaces:
+        weighted["evidence_linkage_penalty"] = -8.0
+    elif len(surfaces) == 1:
+        weighted["evidence_linkage_bonus"] = 0.5
+    else:
+        weighted["evidence_linkage_bonus"] = min(2.0, 0.5 * len(surfaces))
+    provenance = _provenance_counts(surfaces)
+    weighted["source_quality"] = (
+        1.5 * provenance["observed_runtime"]
+        - 0.5 * provenance["protocol_derived"]
+        - 3.0 * provenance["low_quality"]
+    )
+    evidence_blob = f"{finding.title} {' '.join(finding.evidence)}"
+    if _is_low_quality_source(evidence_blob) and all(_is_low_quality_source(surface.file) for surface in surfaces):
+        weighted["source_quality_penalty"] = -10.0
+    return round(sum(weighted.values()), 2), weighted
+
+
+def _overview_line(scan: ScanResult) -> str:
+    repo_type = "web/service-facing" if scan.routes else "non-web"
+    return (
+        f"- AttackMap reviewed a {repo_type} codebase with {len(scan.routes)} inferred entry point"
+        f"{'s' if len(scan.routes) != 1 else ''} across {scan.files_scanned} scanned files."
+    )
+
+
+def _strengths(scan: ScanResult, attack_surfaces: list[AttackSurface]) -> list[str]:
+    items: list[tuple[float, str]] = []
+    auth_hints = sorted({hint.hint for hint in scan.auth_hints})
+    secure_surfaces = [surface for surface in attack_surfaces if surface.auth_signals]
+    internal_surfaces = [surface for surface in attack_surfaces if surface.exposure == "internal"]
+    low_risk_surfaces = [surface for surface in attack_surfaces if surface.risk == "low"]
+
+    if auth_hints:
+        items.append(
+            (
+                9.0 + min(len(auth_hints), 5),
+                f"- Authentication/identity indicators are present in code signals ({', '.join(auth_hints[:6])}).",
+            )
+        )
+    if secure_surfaces:
+        items.append(
+            (
+                8.0 + min(len(secure_surfaces), 5),
+                f"- {len(secure_surfaces)} route(s) include nearby auth indicators, which may support defensive enforcement if validated at runtime.",
+            )
+        )
+    if internal_surfaces:
+        items.append((7.0, f"- {len(internal_surfaces)} route(s) were inferred as internal-only, reducing direct internet exposure if correctly segmented."))
+    if low_risk_surfaces:
+        items.append((6.0, f"- {len(low_risk_surfaces)} route(s) were classified as low-risk operational surfaces."))
+    if scan.secret_hints:
+        items.append((6.5, "- Secret usage appears environment-driven rather than hardcoded literals."))
+    else:
+        items.append((5.0, "- No obvious secret-like environment references were detected in executable paths."))
+    if not items:
+        return ["- No strong defensive indicators were detected heuristically; manual review is still recommended."]
+    return [line for _, line in sorted(items, key=lambda item: item[0], reverse=True)[:4]]
+
+
+def _weaknesses(attack_surfaces: list[AttackSurface], findings: list[Finding], attack_paths: list[AttackPath]) -> list[str]:
+    items: list[str] = []
+    finding_scores = sorted(
+        (
+            (_finding_priority(finding, attack_surfaces, attack_paths), finding)
+            for finding in findings
+        ),
+        key=lambda item: (item[0][0], -_severity_rank(item[1].severity)),
+        reverse=True,
+    )
+    for (score, factors), finding in finding_scores[:3]:
+        related_surfaces = _related_surfaces_for_finding(finding, attack_surfaces)
+        items.append(f"- [{finding.severity.upper()} | score {score:.1f}] {finding.title}")
+        items.append(f"- Reason: {_top_score_reasons(factors, top_n=3)}")
+        items.append(f"- Provenance: {_provenance_breakdown(related_surfaces)}")
+
+    hotspot_surfaces = [surface for surface in attack_surfaces if not _is_low_quality_source(surface.file)]
+    if not hotspot_surfaces:
+        hotspot_surfaces = list(attack_surfaces)
+
+    hotspot_scores = sorted(
+        (
+            (_surface_priority(surface), surface)
+            for surface in hotspot_surfaces
+        ),
+        key=lambda item: (item[0][0], -_surface_risk_rank(item[1].risk), item[1].route),
+        reverse=True,
+    )
+    # Avoid duplicate hotspot rows for the same logical surface (e.g., ANY/SUBSCRIBE variants in one lexicon file).
+    deduped_hotspots: list[tuple[tuple[float, dict[str, float]], AttackSurface]] = []
+    seen_hotspot_keys: set[tuple[str, str, str, str]] = set()
+    for scored_surface in hotspot_scores:
+        (_score, _factors), surface = scored_surface
+        key = (surface.route, surface.file, surface.category, surface.exposure)
+        if key in seen_hotspot_keys:
+            continue
+        seen_hotspot_keys.add(key)
+        deduped_hotspots.append(scored_surface)
+        if len(deduped_hotspots) >= 2:
+            break
+
+    for (score, factors), surface in deduped_hotspots:
+        items.append(
+            f"- Hotspot [score {score:.1f}]: {surface.method} {surface.route} ({surface.location()}) -> {surface.category} / {surface.risk}"
+        )
+        items.append(f"- Reason: {_top_score_reasons(factors, top_n=2)}")
+        items.append(f"- Provenance: {_provenance_breakdown([surface])}")
+    return items[:12]
+
+
+def _evidence_chains(attack_paths: list[AttackPath]) -> list[str]:
+    if not attack_paths:
+        return ["- No explicit attack path was generated from current signals."]
+    lines: list[str] = []
+    scored_paths = sorted(
+        ((_path_priority(path), path) for path in attack_paths),
+        key=lambda item: item[0][0],
+        reverse=True,
+    )
+    for (score, factors), path in scored_paths[:3]:
+        lines.append(f"- [score {score:.1f}] {path.name}: {path.impact}")
+        lines.append(f"- Reason: {_top_score_reasons(factors, top_n=3)}")
+        if path.steps:
+            lines.append(f"- Key step: {path.steps[0]}")
+    return lines
+
+
+def _recommendations(findings: list[Finding], attack_paths: list[AttackPath], attack_surfaces: list[AttackSurface]) -> list[str]:
+    recommendations: list[str] = []
+    low_quality_recommendations: list[str] = []
+    scored_findings = sorted(
+        (
+            (_finding_priority(finding, attack_surfaces, attack_paths), finding)
+            for finding in findings
+        ),
+        key=lambda item: item[0][0],
+        reverse=True,
+    )
+    for (_score, _factors), finding in scored_findings:
+        if not finding.mitigation:
+            continue
+        related_surfaces = _related_surfaces_for_finding(finding, attack_surfaces)
+        basis = _recommendation_basis_label(related_surfaces)
+        line = f"- [{basis}] {finding.mitigation}"
+        if basis == "low-quality-evidence":
+            if line not in low_quality_recommendations:
+                low_quality_recommendations.append(line)
+            continue
+        if line not in recommendations:
+            recommendations.append(line)
+        if len(recommendations) >= 4:
+            break
+
+    for line in low_quality_recommendations:
+        if len(recommendations) >= 4:
+            break
+        if line not in recommendations:
+            recommendations.append(line)
+
+    high_scored_path = max((_path_priority(path)[0] for path in attack_paths), default=0.0)
+    if attack_paths and len(recommendations) < 5 and high_scored_path >= 22.0:
+        recommendations.append(
+            "- Validate every trust boundary hop in attack paths with explicit authn/authz and least-privilege service credentials."
+        )
+
+    if len(recommendations) < 5:
+        recommendations.append(
+            "- Add targeted tests for highest-risk routes and service edges to prevent regressions in boundary enforcement."
+        )
+    return recommendations[:5]
+
+
+def _render_assets_section(assets: list[Asset]) -> list[str]:
+    if not assets:
+        return ["- No high-value assets identified from code signals."]
+    lines: list[str] = []
+    for asset in assets[:8]:
+        head = f"- **{asset.name}** ({asset.kind}, criticality={asset.criticality})"
+        evidence_inline = "; ".join(asset.evidence[:3]) if asset.evidence else ""
+        if evidence_inline:
+            head = f"{head} — {evidence_inline}"
+        lines.append(head)
+    if len(assets) > 8:
+        lines.append(f"- …and {len(assets) - 8} more.")
+    return lines
+
+
+def _render_controls_section(controls: list[Control]) -> list[str]:
+    if not controls:
+        return ["- No defensive controls observed from code signals."]
+    present = [c for c in controls if c.strength != "absent"]
+    absent = [c for c in controls if c.strength == "absent"]
+    lines: list[str] = []
+    if present:
+        lines.append("**Observed:**")
+        for control in present[:8]:
+            placement = f" — placements: {len(control.placements)}" if control.placements else ""
+            lines.append(f"- `{control.kind}` · {control.name} (strength={control.strength}){placement}")
+        if len(present) > 8:
+            lines.append(f"- …and {len(present) - 8} more.")
+    if absent:
+        lines.append("")
+        lines.append("**Expected but not observed (control absences):**")
+        for control in absent[:6]:
+            note = f" — {control.notes}" if control.notes else ""
+            lines.append(f"- `{control.kind}` · {control.name}{note}")
+    return lines
+
+
+def _render_notable_observations(insights: list[Insight]) -> list[str]:
+    if not insights:
+        return ["- No notable cross-cutting observations were generated for this scan."]
+    lines: list[str] = []
+    for insight in insights[:6]:
+        head = f"- **[{insight.severity.upper()}]** {insight.title} _(confidence={insight.confidence}, kind={insight.kind})_"
+        lines.append(head)
+        lines.append(f"  - {insight.narrative}")
+        if insight.suggested_action:
+            lines.append(f"  - _Suggested action:_ {insight.suggested_action}")
+        if insight.attack_techniques:
+            techniques_text = ", ".join(
+                f"`{t.technique_id}` {t.name}" for t in insight.attack_techniques[:4]
+            )
+            lines.append(f"  - _ATT&CK:_ {techniques_text}")
+        if insight.evidence:
+            lines.append(f"  - _Evidence:_ {'; '.join(insight.evidence[:3])}")
+    if len(insights) > 6:
+        lines.append(f"- …and {len(insights) - 6} more notable observations in JSON output.")
+    return lines
+
+
+def _render_detection_opportunities(opportunities: list[DetectionOpportunity]) -> list[str]:
+    if not opportunities:
+        return ["- No detection opportunities were generated for this scan."]
+    lines: list[str] = []
+    for opp in opportunities[:5]:
+        lines.append(f"- **{opp.title}** _(signal={opp.signal_kind})_")
+        lines.append(f"  - _Why:_ {opp.rationale}")
+        lines.append(f"  - _Suggested rule:_ {opp.suggested_rule}")
+        if opp.attack_techniques:
+            techniques_text = ", ".join(
+                f"`{t.technique_id}`" for t in opp.attack_techniques[:4]
+            )
+            lines.append(f"  - _ATT&CK:_ {techniques_text}")
+    if len(opportunities) > 5:
+        lines.append(f"- …and {len(opportunities) - 5} more detection opportunities in JSON output.")
+    return lines
+
+
+def _render_attack_techniques_summary(overlay: SecurityOverlay) -> list[str]:
+    by_id: dict[str, tuple[str, str]] = {}
+    for insight in overlay.insights:
+        for t in insight.attack_techniques:
+            by_id.setdefault(t.technique_id, (t.name, t.tactic))
+    for finding in overlay.findings:
+        for t in finding.attack_techniques:
+            by_id.setdefault(t.technique_id, (t.name, t.tactic))
+    if not by_id:
+        return ["- No ATT&CK techniques were mapped from the current findings/insights."]
+    lines: list[str] = []
+    for tid in sorted(by_id):
+        name, tactic = by_id[tid]
+        lines.append(f"- `{tid}` **{name}** — _{tactic}_")
+    return lines
+
+
+def render_defensive_review(
+    scan: ScanResult,
+    attack_surfaces: list[AttackSurface],
+    findings: list[Finding],
+    attack_paths: list[AttackPath],
+) -> str:
+    observed_runtime_public, protocol_derived, internal_only, low_quality = _entrypoint_counts(attack_surfaces)
+    overlay = build_security_overlay(scan, attack_surfaces, findings, attack_paths)
+    lines = [
+        "# Defensive Review",
+        "",
+        "## System Overview",
+        _overview_line(scan),
+        f"- Raw inferred entry points: {len(scan.routes)}",
+        f"- Observed runtime/public surfaces: {observed_runtime_public}",
+        f"- Protocol/lexicon-derived surfaces (inferred): {protocol_derived}",
+        f"- Internal-only surfaces: {internal_only}",
+        f"- Test/example/mocked surfaces (down-weighted): {low_quality}",
+        f"- Languages: {', '.join(scan.languages) if scan.languages else 'none'}",
+        f"- Datastores: {', '.join(sorted({db.kind for db in scan.databases})) if scan.databases else 'none'}",
+        f"- External dependencies observed: {len(scan.external_calls)}",
+        f"- Identified assets: {len(overlay.assets)} | controls observed: "
+        f"{sum(1 for c in overlay.controls if c.strength != 'absent')} | "
+        f"control absences: {sum(1 for c in overlay.controls if c.strength == 'absent')}",
+        "",
+        "## Notable Observations",
+        *_render_notable_observations(overlay.insights),
+        "",
+        "## ATT&CK Techniques Mapped",
+        *_render_attack_techniques_summary(overlay),
+        "",
+        "## Detection Opportunities",
+        *_render_detection_opportunities(overlay.detection_opportunities),
+        "",
+        "## Asset Inventory",
+        *_render_assets_section(overlay.assets),
+        "",
+        "## Defensive Controls",
+        *_render_controls_section(overlay.controls),
+        "",
+        "## Strengths",
+        *_strengths(scan, attack_surfaces),
+        "",
+        "## Weaknesses / Risk Hotspots",
+        *_weaknesses(attack_surfaces, findings, attack_paths),
+        "",
+        "## Key Evidence Chains",
+        *_evidence_chains(attack_paths),
+        "",
+        "## Recommendations",
+        *_recommendations(findings, attack_paths, attack_surfaces),
+        "",
+        "## Analyst Notes",
+        "- This defensive review is heuristic and intended as an engineering triage starting point.",
+        "- Validate top risks with repository-specific architecture context and runtime controls.",
+        "- Notable observations are cross-cutting hypotheses; treat them as triage prompts, not verdicts.",
+    ]
+    return "\n".join(lines)