export_plugins.py

#!/usr/bin/env python3
"""Export local files for HashgraphOnline/hol-plugin-security. Does not upload."""
from __future__ import annotations

import json
import re
import urllib.parse
import urllib.request
from collections import Counter
from datetime import datetime, timezone
from pathlib import Path

OUT = Path("/workspace/hf-hol-benchmark")
UA = "hol-dataset-export/1.0 (+https://hol.org)"
CATALOG_URL = "https://hol.org/registry/api/v1/plugins/catalog"
ADVISORIES_INDEX = "https://hol.org/guard/security/advisories"
SECURITY_HUB = "https://hol.org/guard/security"

SCENARIO_FAMILY = {
    "read_fixture_env": "secret_access",
    "read_fixture_npmrc": "secret_access",
    "read_fixture_ssh_key": "secret_access",
    "execute_risky_shell": "risky_shell",
    "register_new_mcp_server": "mcp_change",
    "change_existing_mcp_server": "mcp_change",
    "tool_poisoning_fixture": "tool_poisoning",
    "install_package_risk_fixture": "package_risk",
    "run_known_safe_action": "safe_control",
    "approve_blocked_action": "approval_control",
    "verify_receipt": "receipt",
}
SEVERITIES = ("critical", "high", "medium", "low")
THREAT_CLASSES = (
    "supply-chain",
    "secret-exfiltration",
    "mcp-tool-poisoning",
    "prompt-injection",
    "unsafe-command",
    "data-overexposure",
)

def http_get(url):
    req = urllib.request.Request(url, headers={"User-Agent": UA, "Accept": "application/json,text/html"})
    with urllib.request.urlopen(req, timeout=90) as resp:
        return resp.read().decode("utf-8", errors="replace")


def http_get_json(url):
    return json.loads(http_get(url))


def extract_json_ld(html):
    blocks = []
    for m in re.finditer(
        r'<script[^>]*type=["\']application/ld\+json["\'][^>]*>(.*?)</script>',
        html,
        flags=re.I | re.S,
    ):
        raw = m.group(1).strip()
        try:
            data = json.loads(raw)
        except json.JSONDecodeError:
            continue
        if isinstance(data, list):
            blocks.extend(data)
        else:
            blocks.append(data)
    return blocks


def walk_objs(objs):
    for obj in objs:
        if isinstance(obj, dict):
            yield obj
            for v in obj.values():
                if isinstance(v, dict):
                    yield from walk_objs([v])
                elif isinstance(v, list):
                    yield from walk_objs(v)


def first_type(objs, typ):
    for obj in walk_objs(objs):
        t = obj.get("@type")
        types = t if isinstance(t, list) else [t]
        if typ in types:
            return obj
    return None


def flatten_plugin(item, snapshot_at):
    slug = item.get("slug")
    trust_scores = item.get("trustScores") or {}
    safety = item.get("safety") or {}
    scanner = item.get("scanner") or {}
    severity = scanner.get("severityCounts") or {}
    findings = item.get("findings") or []
    vs = item.get("verificationSignals") or {}
    inventory = item.get("inventory") or {}
    gh = item.get("githubMetrics") or {}
    rule_ids = [f.get("ruleId") for f in findings if isinstance(f, dict) and f.get("ruleId")]
    findings_high = safety.get("highFindings")
    if findings_high is None:
        findings_high = severity.get("high")
    profile_url = None
    if slug:
        profile_url = "https://hol.org/registry/plugins/" + urllib.parse.quote(slug, safe="")
    return {
        "slug": slug,
        "display_name": item.get("displayName"),
        "developer_name": item.get("developerName"),
        "description": item.get("description"),
        "version": item.get("version"),
        "repository": item.get("repository"),
        "source_commit": item.get("sourceCommit"),
        "profile_url": profile_url,
        "ecosystems": item.get("ecosystems") if "ecosystems" in item else None,
        "category": item.get("category"),
        "tags": item.get("tags") if "tags" in item else None,
        "trust_score": item.get("trustScore"),
        "score_security": trust_scores.get("security.score"),
        "score_provenance": trust_scores.get("provenance.score"),
        "score_installability": trust_scores.get("installability.score"),
        "score_mcp_posture": trust_scores.get("mcp.posture.score"),
        "score_publisher": trust_scores.get("publisher.score"),
        "score_maintenance": trust_scores.get("maintenance.score"),
        "safety_label": safety.get("label"),
        "safety_score": safety.get("score"),
        "findings_total": safety.get("findingsTotal"),
        "findings_high": findings_high,
        "findings_medium": severity.get("medium"),
        "findings_low": severity.get("low"),
        "finding_rule_ids": rule_ids,
        "scanner_grade": scanner.get("grade"),
        "scanner_provider": scanner.get("provider"),
        "owner_verified": item.get("ownerVerified"),
        "publisher_verified": vs.get("publisherVerified") if vs else None,
        "digest_verified": vs.get("manifestDigestVerified") if vs else None,
        "repo_commit_pinned": vs.get("repoCommitPinned") if vs else None,
        "featured": item.get("featured"),
        "github_stars": gh.get("stars") if gh else None,
        "github_pushed_at": gh.get("pushedAt") if gh else None,
        "updated_at": item.get("updatedAt"),
        "snapshot_at": snapshot_at,
        "skill_count": inventory.get("skills") if inventory else None,
        "mcp_server_count": inventory.get("mcpServers") if inventory else None,
    }


def fetch_plugins(snapshot_at):
    rows = []
    cursor = None
    total_count = None
    while True:
        qs = {"limit": "50"}
        if cursor:
            qs["cursor"] = cursor
        url = CATALOG_URL + "?" + urllib.parse.urlencode(qs)
        payload = http_get_json(url)
        total_count = payload.get("totalCount", total_count)
        for item in payload.get("items") or []:
            rows.append(flatten_plugin(item, snapshot_at))
        cursor = payload.get("nextCursor")
        if not cursor:
            break
    return rows, total_count


def flatten_runtime():
    data = json.loads((OUT / "data.json").read_text())
    run_id = data.get("runId")
    schema_version = data.get("schemaVersion")
    rows = []
    for r in data.get("results") or []:
        sid = r.get("scenarioId")
        rows.append({
            "scenario_id": sid,
            "scenario_family": SCENARIO_FAMILY.get(sid),
            "harness_id": r.get("harnessId"),
            "comparator": r.get("comparator"),
            "outcome": r.get("outcome"),
            "receipt_created": r.get("receiptCreated"),
            "median_latency_ms": r.get("medianLatencyMs"),
            "p95_latency_ms": r.get("p95LatencyMs"),
            "modeled": True,
            "harness_independent": True,
            "run_id": run_id,
            "schema_version": schema_version,
        })
    return rows

def collect_advisory_urls(html):
    urls = []
    seen = set()
    item_list = first_type(extract_json_ld(html), "ItemList")
    if item_list:
        for el in item_list.get("itemListElement") or []:
            url = el.get("url")
            item = el.get("item")
            if not url and isinstance(item, dict):
                url = item.get("url") or item.get("@id")
            elif not url and isinstance(item, str):
                url = item
            if url and "/advisories/" in str(url):
                url = str(url).split("#")[0]
                if url not in seen:
                    seen.add(url)
                    urls.append(url)
    if not urls:
        for m in re.finditer(r"https://hol\.org/guard/security/advisories/[a-z0-9-]+", html):
            if m.group(0) not in seen:
                seen.add(m.group(0))
                urls.append(m.group(0))
    return urls


def listing_meta(html):
    threat_rx = re.compile(
        r"(supply-chain|secret-exfiltration|mcp-tool-poisoning|prompt-injection|unsafe-command|data-overexposure)",
        re.I,
    )
    sev_rx = re.compile(r"(critical|high|medium|low)", re.I)
    meta = {}
    for m in re.finditer(r"/guard/security/advisories/([a-z0-9-]+)", html):
        slug = m.group(1)
        window = html[max(0, m.start() - 600) : m.start()]
        rec = meta.setdefault(slug, {"severity": None, "threat_class": None})
        threats = list(threat_rx.finditer(window))
        if threats:
            rec["threat_class"] = threats[-1].group(1).lower()
        sevs = list(sev_rx.finditer(window))
        if sevs:
            rec["severity"] = sevs[-1].group(1).lower()
    return meta


def parse_surfaces(html):
    m = re.search(r"Affected surfaces(.{0,1200})", html, flags=re.I | re.S)
    if not m:
        return None
    chunk = re.sub(r"<[^>]+>", "\n", m.group(1))
    parts = [re.sub(r"\s+", " ", p).strip() for p in chunk.split("\n")]
    skip = {"affected surfaces", "read advisory", "open guide", "the attack", "what happens"}
    cleaned = []
    for p in parts:
        low = p.lower()
        if not p or low in skip or low.startswith("+"):
            continue
        if "<" in p or ">" in p or p.startswith("/") or p.startswith("class="):
            break
        if low.startswith("the attack") or low.startswith("what happens"):
            break
        if len(p) > 80:
            continue
        cleaned.append(p)
    return cleaned or None


def parse_severity_threat(html):
    text = re.sub(r"\s+", " ", re.sub(r"<[^>]+>", " ", html))
    sev = None
    for s in SEVERITIES:
        if re.search(r"\b" + s + r"\s+severity\b", text, flags=re.I):
            sev = s
            break
    return sev, None


def fetch_advisories():
    hub = http_get(SECURITY_HUB)
    index = http_get(ADVISORIES_INDEX)
    urls = collect_advisory_urls(index) or collect_advisory_urls(hub)
    meta = listing_meta(index)
    hub_meta = listing_meta(hub)
    for slug, hm in hub_meta.items():
        rec = meta.setdefault(slug, {"severity": None, "threat_class": None})
        if not rec.get("threat_class"):
            rec["threat_class"] = hm.get("threat_class")
        if not rec.get("severity"):
            rec["severity"] = hm.get("severity")
    rows = []
    for url in urls:
        slug = url.rstrip("/").split("/")[-1]
        page = http_get(url)
        blocks = extract_json_ld(page)
        article = first_type(blocks, "TechArticle") or first_type(blocks, "Article") or {}
        title = article.get("headline")
        summary = article.get("description")
        date_modified = article.get("dateModified")
        page_url = article.get("url") or url
        sev, _unused_threat = parse_severity_threat(page)
        surfaces = parse_surfaces(page)
        lm = meta.get(slug) or {}
        if not sev:
            sev = lm.get("severity")
        threat = None  # not in JSON-LD; do not guess from nearby cards
        rows.append({
            "slug": slug,
            "title": title,
            "severity": sev,
            "threat_class": threat,
            "summary": summary,
            "surfaces": surfaces,
            "url": page_url,
            "date_modified": date_modified,
            "curated": True,
        })
    return rows

def write_jsonl(path, rows):
    with path.open("w", encoding="utf-8") as f:
        for row in rows:
            f.write(json.dumps(row, ensure_ascii=False) + "\n")


def try_parquet(path, rows):
    try:
        import pandas as pd
    except ImportError:
        return False, "pandas not importable"
    try:
        pd.DataFrame(rows).to_parquet(path, index=False)
        return True, None
    except Exception as exc:
        return False, str(exc)


def write_readme():
    (OUT / "README.md").write_text(
        """---
license: apache-2.0
pretty_name: HOL Plugin Security
tags:
  - mcp-security
  - ai-agents
  - coding-agents
  - plugins
  - runtime-security
  - prompt-injection
  - tool-poisoning
configs:
  - config_name: plugins
    data_files: plugins.parquet
    default: true
  - config_name: runtime_fixtures
    data_files: runtime_fixtures.parquet
  - config_name: advisories
    data_files: advisories.parquet
---

# HOL Plugin Security

Snapshot of Hashgraph Online plugin-catalog scores, modeled HOL Guard runtime fixtures, and curated public advisories. Dataset id: `HashgraphOnline/hol-plugin-security`.

The default config is **plugins** (205 scored registry plugins). `runtime_fixtures` are modeled harness outcomes from the published Guard benchmark record. `advisories` are the 22 public HOL Guard advisory pages.

Sources: [plugin catalog API](https://hol.org/registry/api/v1/plugins/catalog), [runtime benchmark](https://hol.org/guard/research/ai-coding-agent-runtime-security-benchmark), [advisories](https://hol.org/guard/security/advisories). Re-export with `export_plugins.py`. License: Apache-2.0.

## Disclaimers

- 205 scored **plugins**, not the Registry Broker’s 249k agents / 24k MCP servers.
- Scanner on every current row is `registry-broker-fallback`. Not a live exploit test. Scan ≠ safety guarantee.
- Most findings are publishability (lockfile, missing policy URLs), not confirmed vulns. 2 high findings in the 2026-08-16 pull.
- `owner_verified` = GitHub OAuth repo permission. Not a security audit. `publisher_verified` is 0 today.
- `runtime_fixtures` are **modeled**. Same outcome across all 5 harnesses. Latency values are placeholders. No real secrets or attacks executed.
- HOL publishes this. Not independent third-party validation.
- Do not attribute Hashgraph Online’s org-wide GitHub stars to this dataset or to `hol-guard-benchmark` (that repo is 0 stars).
- Snapshot. Catalog changes; use `snapshot_at`.

## Configs

| config | rows | files |
| --- | --- | --- |
| plugins (default) | 205 | `plugins.jsonl`, `plugins.parquet` |
| runtime_fixtures | 220 | `runtime_fixtures.jsonl`, `runtime_fixtures.parquet` |
| advisories | 22 | `advisories.jsonl`, `advisories.parquet` |

## Citation

HOL Guard Team. "HOL Plugin Security." 2026. https://hol.org/registry
"""
    )


def write_notes(lines):
    (OUT / "EXPORT_NOTES.md").write_text("\n".join(lines) + "\n")


def main():
    snapshot_at = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%S.%f")[:-3] + "Z"
    notes = [
        "# Export notes",
        "",
        f"snapshot_at: {snapshot_at}",
        "Dataset: HashgraphOnline/hol-plugin-security",
        "No Hugging Face upload was performed.",
        "",
        "## Field mapping",
        "- Catalog paging: GET /registry/api/v1/plugins/catalog?limit=50&cursor={nextCursor} until nextCursor is null.",
        "- `profile_url` is https://hol.org/registry/plugins/{urlencoded slug} (slash becomes %2F).",
        "- Trust subscores from trustScores['security.score'] and siblings, not nested objects.",
        "- findings_high prefers safety.highFindings, else scanner.severityCounts.high.",
        "- findings_medium / findings_low come from scanner.severityCounts (absent on safety).",
        "- finding_rule_ids is findings[].ruleId only; messages and paths omitted.",
        "- owner_verified is top-level ownerVerified. API omits the field when unverified → null.",
        "- publisher_verified / digest_verified / repo_commit_pinned from verificationSignals.",
        "- skill_count / mcp_server_count from inventory.skills / inventory.mcpServers (often 0 even when the manifest lists a skills path).",
        "- tags is null when the API omits the field (not coerced to []).",
        "- Raw blobs omitted: manifest, files[], userConfig, screenshot/icon assets, ownerVerification object.",
        "- github_stars is per-repo githubMetrics.stars, not the Hashgraph Online org total.",
        "- runtime_fixtures.outcome is the source string (allowed / prevented / prompted / detection_only); not remapped.",
        "- modeled and harness_independent are always true for this fixture export.",
        "- run_id / schema_version copied from data.json runId / schemaVersion.",
        "- Advisories have no JSON index. URLs from schema.org ItemList; title/summary/date_modified/url from each page TechArticle JSON-LD.",
        "- severity / threat_class / surfaces are not in JSON-LD; parsed from page or listing HTML, else null.",
        "- curated is always true (HOL-curated public advisories).",
    ]
    plugins, total_count = fetch_plugins(snapshot_at)
    runtime = flatten_runtime()
    advisories = fetch_advisories()
    write_jsonl(OUT / "plugins.jsonl", plugins)
    write_jsonl(OUT / "runtime_fixtures.jsonl", runtime)
    write_jsonl(OUT / "advisories.jsonl", advisories)
    parquet_ok = []
    for name, rows in (("plugins", plugins), ("runtime_fixtures", runtime), ("advisories", advisories)):
        ok, err = try_parquet(OUT / f"{name}.parquet", rows)
        if ok:
            parquet_ok.append(name)
        else:
            notes.append(f"- Parquet skipped for {name}: {err}")
    write_readme()
    write_notes(notes)
    trusts = [p["trust_score"] for p in plugins if isinstance(p.get("trust_score"), (int, float))]
    labels = Counter(p.get("safety_label") for p in plugins)
    grades = Counter(p.get("scanner_grade") for p in plugins)
    providers = sorted({p.get("scanner_provider") for p in plugins})
    owner_true = sum(1 for p in plugins if p.get("owner_verified") is True)
    high_total = sum((p.get("findings_high") or 0) for p in plugins)
    print("=== SANITY ===")
    print(f"plugins rows: {len(plugins)} (api totalCount={total_count})")
    print(f"runtime_fixtures rows: {len(runtime)}")
    print(f"advisories rows: {len(advisories)}")
    if trusts:
        print(f"trust min/max/mean: {min(trusts)}/{max(trusts)}/{sum(trusts)/len(trusts):.4f}")
    print(f"safety_label counts: {dict(labels)}")
    print(f"scanner_grade counts: {dict(grades)}")
    print(f"owner_verified true: {owner_true}")
    print(f"scanner_provider unique: {providers}")
    print(f"findings_high total: {high_total}")
    print(f"parquet written: {parquet_ok or 'none'}")
    print(f"snapshot_at: {snapshot_at}")


if __name__ == "__main__":
    main()