/**
 * Analysis for research report EFF-RR-2026-003: vulnerabilities in AI application and agent tooling.
 *
 *   node research/ai-agents-2026/analyse.mjs <data-dir> <out-dir>
 *
 * <data-dir> holds nvd-2023.json.gz ... nvd-2026.json.gz (NVD CVE 2.0 yearly feeds) and kev.json
 * (CISA KEV catalogue), the same inputs as research/epss-2026. See README.md.
 *
 * Outputs: results.json (every number in the report) and ai-tooling-cves.csv (one row per CVE).
 */
import { createReadStream, readFileSync, writeFileSync, mkdirSync } from 'node:fs';
import { createGunzip } from 'node:zlib';
import { join } from 'node:path';

const [dir, out] = process.argv.slice(2);
if (!out) { console.error('Usage: node analyse.mjs <data-dir> <out-dir>'); process.exit(1); }
const YEARS = [2023, 2024, 2025, 2026];

/* ---------- Product rules ---------- */

// Matched against "vendor | product" pairs from the CNA's affected-product data and NVD's CPE data,
// lower-cased. Order matters: the first rule that matches any pair in a CVE assigns its product.
const GROUPS = {
  assistant: 'AI coding assistants and copilots',
  mcp: 'MCP servers and tooling',
  framework: 'Agent frameworks and SDKs',
  builder: 'AI app builders and chat platforms',
  serving: 'Model serving and gateways',
  ml: 'ML platforms and libraries',
};
const PRODUCTS = [
  [/^anthropics? \| claude[-_ ]code/, 'Claude Code', 'assistant'],
  [/^(cursor|anysphere) \| cursor/, 'Cursor', 'assistant'],
  [/^roo ?code(inc)? \| roo[-_ ]code/, 'Roo Code', 'assistant'],
  [/^cline(bot)? \| cline/, 'Cline', 'assistant'],
  [/^google \| gemini[-_ ]cli/, 'Gemini CLI', 'assistant'],
  [/^openai \| codex/, 'OpenAI Codex CLI', 'assistant'],
  [/^(microsoft|github) \| .*copilot/, 'Microsoft and GitHub Copilot', 'assistant'],
  [/^(?!.*wordpress)(?!.*\bwp).*(\| (.*[-_ /])?mcp([-_ ]|$)|mcphub|model[-_ ]context[-_ ]protocol)/, 'MCP servers and tooling', 'mcp'],
  [/langchain|langgraph/, 'LangChain and LangGraph', 'framework'],
  [/llama[-_ ]?index|run-llama/, 'LlamaIndex', 'framework'],
  [/\| (microsoft\/)?autogen/, 'AutoGen', 'framework'],
  [/crewai/, 'CrewAI', 'framework'],
  [/semantic[-_ ]kernel/, 'Semantic Kernel', 'framework'],
  [/significant-gravitas|^agpt \||\| autogpt/, 'AutoGPT', 'framework'],
  [/superagi/, 'SuperAGI', 'framework'],
  [/metagpt/, 'MetaGPT', 'framework'],
  [/agentscope/, 'AgentScope', 'framework'],
  [/praisonai/, 'PraisonAI', 'framework'],
  [/hermes-agent/, 'Hermes Agent', 'framework'],
  [/cowagent/, 'CowAgent', 'framework'],
  [/smolagents/, 'smolagents', 'framework'],
  [/langflow/, 'Langflow', 'builder'],
  [/flowise/, 'Flowise', 'builder'],
  [/^(langgenius|dify) \|/, 'Dify', 'builder'],
  [/^n8n(-io)? \|/, 'n8n', 'builder'],
  [/fastgpt/, 'FastGPT', 'builder'],
  [/anything-?llm/, 'AnythingLLM', 'builder'],
  [/open[-_ ]?webui/, 'Open WebUI', 'builder'],
  [/librechat/, 'LibreChat', 'builder'],
  [/lobe-?chat|^lobehub \|/, 'LobeChat', 'builder'],
  [/lollms/, 'LoLLMs', 'builder'],
  [/chuanhuchatgpt/, 'ChuanhuChatGPT', 'builder'],
  [/gpt_academic/, 'GPT Academic', 'builder'],
  [/nextchat|chatgptnextweb/, 'NextChat', 'builder'],
  [/db-gpt/, 'DB-GPT', 'builder'],
  [/privategpt/, 'PrivateGPT', 'builder'],
  [/text-generation-webui/, 'Text Generation WebUI', 'builder'],
  [/vllm/, 'vLLM', 'serving'],
  [/sglang/, 'SGLang', 'serving'],
  [/llama\.cpp/, 'llama.cpp', 'serving'],
  [/^ollama \|/, 'Ollama', 'serving'],
  [/litellm/, 'LiteLLM', 'serving'],
  [/localai/, 'LocalAI', 'serving'],
  [/lightllm/, 'LightLLM', 'serving'],
  [/tensorrt-llm/, 'TensorRT-LLM', 'serving'],
  [/triton[_ ]inference[_ ]server/, 'Triton Inference Server', 'serving'],
  [/xinference/, 'Xinference', 'serving'],
  [/gpustack/, 'GPUStack', 'serving'],
  [/bentoml/, 'BentoML', 'serving'],
  [/mlflow/, 'MLflow', 'ml'],
  [/^h2o(ai)? \| (h2oai\/)?h2o(-3)?$/, 'H2O-3', 'ml'],
  [/^(ray-project|anyscale) \|/, 'Ray', 'ml'],
  [/kubeflow/, 'Kubeflow', 'ml'],
  [/^(huggingface|hugging face) \| (huggingface\/)?transformers/, 'Hugging Face Transformers', 'ml'],
  [/gradio/, 'Gradio', 'ml'],
  [/clearml/, 'ClearML', 'ml'],
];

// Weakness families, the same as report EFF-RR-2026-001, plus prompt injection (CWE-1427).
const FAMILIES = {
  prompt: ['Prompt injection', [1427]],
  memory: ['Memory safety', [119, 120, 121, 122, 124, 125, 126, 129, 131, 190, 191, 401, 415, 416, 476, 680, 704, 763, 787, 822, 823, 824, 843]],
  auth: ['Authentication and access control', [250, 266, 267, 269, 276, 284, 285, 287, 288, 290, 294, 302, 303, 304, 305, 306, 307, 345, 346, 347, 425, 521, 522, 639, 640, 648, 732, 798, 862, 863, 1188, 1390]],
  injection: ['Injection and code execution', [74, 77, 78, 88, 89, 91, 93, 94, 95, 96, 611, 917, 1236, 1321, 1336]],
  xss: ['Cross-site scripting', [79, 80]],
  file: ['Path traversal and file handling', [22, 23, 24, 29, 35, 36, 59, 61, 73, 434, 552]],
  deser: ['Unsafe deserialisation', [502]],
  forgery: ['Request forgery (SSRF, CSRF)', [352, 918]],
  resource: ['Resource exhaustion', [400, 770, 674, 1333, 834, 835]],
  exposure: ['Information exposure', [200, 201, 209, 532, 538]],
};
const familyOf = new Map(Object.entries(FAMILIES).flatMap(([id, [, c]]) => c.map((n) => [`CWE-${n}`, id])));
const FAMILY_LABEL = { ...Object.fromEntries(Object.entries(FAMILIES).map(([id, [l]]) => [id, l])), other: 'Other specified weakness', none: 'No weakness recorded' };

/* ---------- Reading ---------- */

async function* vulnerabilities(file) {
  const stream = createReadStream(file).pipe(createGunzip());
  stream.setEncoding('utf8');
  let buf = '', started = false, depth = 0, inStr = false, esc = false, start = 0, pos = 0;
  for await (const chunk of stream) {
    buf += chunk;
    if (!started) { const i = buf.indexOf('"vulnerabilities"'); const j = i < 0 ? -1 : buf.indexOf('[', i); if (j < 0) continue; buf = buf.slice(j + 1); started = true; pos = 0; }
    for (; pos < buf.length; pos++) {
      const c = buf[pos];
      if (inStr) { if (esc) esc = false; else if (c === '\\') esc = true; else if (c === '"') inStr = false; continue; }
      if (c === '"') inStr = true;
      else if (c === '{') { if (depth++ === 0) start = pos; }
      else if (c === '}' && --depth === 0) { yield JSON.parse(buf.slice(start, pos + 1)).cve; buf = buf.slice(pos + 1); pos = -1; }
      else if (c === ']' && depth === 0) return;
    }
    if (depth === 0) { buf = ''; pos = 0; }
  }
}

/**
 * The product a CVE is about: the first product in the CNA's affected list, or, if the CNA gave none,
 * NVD's CPE products. Using only the first CNA product stops multi-product incidents (for example a
 * compromised build tool listing every downstream package) being counted as AI tooling flaws.
 */
function primaryPairs(c) {
  const d = (c.affected || []).flatMap((a) => a.affectedData || [])[0];
  if (d && (d.vendor || d.product)) return [`${(d.vendor || '').toLowerCase().trim()} | ${(d.product || '').toLowerCase().trim()}`];
  const set = new Set();
  const walk = (n) => { for (const m of n.cpeMatch || []) { const p = m.criteria.split(':'); set.add(`${p[3]} | ${p[4]}`); } for (const k of n.children || []) walk(k); };
  for (const cfg of c.configurations || []) for (const n of cfg.nodes || []) walk(n);
  return [...set];
}
function product(ps) {
  for (const [re, name, group] of PRODUCTS) if (ps.some((p) => re.test(p))) return { name, group };
  return null;
}
function cvss3(metrics = {}) {
  const all = [...(metrics.cvssMetricV31 || []), ...(metrics.cvssMetricV30 || [])];
  const cna = all.find((x) => x.type === 'Secondary')?.cvssData.baseScore ?? null;
  const nvd = all.find((x) => x.type === 'Primary')?.cvssData.baseScore ?? null;
  return cna ?? nvd;
}
const band = (s) => (s === null ? 'none' : s >= 9 ? 'critical' : s >= 7 ? 'high' : s >= 4 ? 'medium' : s > 0 ? 'low' : 'none');
function cwes(c) {
  const ids = [];
  for (const w of c.weaknesses || []) for (const d of w.description || []) if (/^CWE-\d+$/.test(d.value) && !ids.includes(d.value)) ids.push(d.value);
  return ids;
}
function family(ids) {
  for (const id of ids) if (familyOf.has(id)) return familyOf.get(id);
  return ids.length ? 'other' : 'none';
}

/* ---------- Run ---------- */

const kevDoc = JSON.parse(readFileSync(join(dir, 'kev.json'), 'utf8'));
const kev = new Map(kevDoc.vulnerabilities.map((v) => [v.cveID, v.dateAdded]));

const rows = [];
let allCves = 0, promptMentionsAll = 0, cwe1427All = 0;
const allByYear = {};
for (const y of YEARS) {
  for await (const c of vulnerabilities(join(dir, `nvd-${y}.json.gz`))) {
    if (c.vulnStatus === 'Rejected') continue;
    allCves++;
    const py = c.published.slice(0, 4);
    allByYear[py] = (allByYear[py] || 0) + 1;
    const desc = (c.descriptions || []).find((d) => d.lang === 'en')?.value || '';
    const mentionsPrompt = /prompt[- ]injection/i.test(desc);
    if (mentionsPrompt) promptMentionsAll++;
    if ((c.weaknesses || []).some((w) => w.description.some((d) => d.value === 'CWE-1427'))) cwe1427All++;
    const p = product(primaryPairs(c));
    if (!p) continue;
    const ids = cwes(c);
    const score = cvss3(c.metrics);
    rows.push({ id: c.id, published: c.published.slice(0, 10), year: py, product: p.name, group: p.group, score, band: band(score), family: family(ids), cwes: ids.join(' '), promptInDescription: mentionsPrompt, kevAdded: kev.get(c.id) || null });
  }
}

/* ---------- Results ---------- */

const pct = (n, d) => (d ? Math.round((n / d) * 1000) / 10 : 0);
const tally = (xs, f) => xs.reduce((m, x) => ((m[f(x)] = (m[f(x)] || 0) + 1), m), {});
const sorted = (o) => Object.entries(o).sort((a, b) => b[1] - a[1]);
const pubYears = ['2023', '2024', '2025', '2026'];

const results = {
  sources: { kevCatalogVersion: kevDoc.catalogVersion, nvdYears: YEARS },
  scope: 'Non-rejected CVEs with 2023-2026 identifiers whose affected-product or CPE data matches a product rule.',
  allCves,
  aiCves: rows.length,
  aiPct: Math.round((rows.length / allCves) * 10000) / 100,
  byPublishedYear: Object.fromEntries(pubYears.map((y) => [y, rows.filter((r) => r.year === y).length])),
  allByPublishedYear: Object.fromEntries(pubYears.map((y) => [y, allByYear[y] || 0])),
  sharePerMilleByYear: Object.fromEntries(pubYears.map((y) => [y, Math.round((rows.filter((r) => r.year === y).length / (allByYear[y] || 1)) * 10000) / 10])),
  byGroup: Object.fromEntries(Object.keys(GROUPS).map((g) => { const s = rows.filter((r) => r.group === g); return [g, { label: GROUPS[g], n: s.length, byYear: Object.fromEntries(pubYears.map((y) => [y, s.filter((r) => r.year === y).length])) }]; })),
  products: sorted(tally(rows, (r) => r.product)).map(([name, n]) => ({ name, n, group: rows.find((r) => r.product === name).group })),
  bands: Object.fromEntries(['critical', 'high', 'medium', 'low', 'none'].map((b) => { const n = rows.filter((r) => r.band === b).length; return [b, { n, pct: pct(n, rows.length) }]; })),
  families: Object.fromEntries(sorted(tally(rows, (r) => r.family)).map(([k, n]) => [k, { label: FAMILY_LABEL[k], n, pct: pct(n, rows.length) }])),
  familiesByGroup: Object.fromEntries(Object.keys(GROUPS).map((g) => { const s = rows.filter((r) => r.group === g); return [g, Object.fromEntries(sorted(tally(s, (r) => r.family)).map(([k, n]) => [k, { n, pct: pct(n, s.length) }]))]; })),
  topCwes: sorted(tally(rows.flatMap((r) => r.cwes.split(' ').filter(Boolean).map((c) => ({ c }))), (x) => x.c)).slice(0, 12),
  promptInjection: {
    cwe1427All, descriptionMentionsAll: promptMentionsAll,
    descriptionMentionsAi: rows.filter((r) => r.promptInDescription).length,
    aiMentionFamilies: Object.fromEntries(sorted(tally(rows.filter((r) => r.promptInDescription), (r) => r.family))),
    aiMentionProducts: sorted(tally(rows.filter((r) => r.promptInDescription), (r) => r.product)).slice(0, 10),
  },
  kev: rows.filter((r) => r.kevAdded).map((r) => ({ id: r.id, product: r.product, published: r.published, kevAdded: r.kevAdded, band: r.band, family: r.family })),
};

mkdirSync(out, { recursive: true });
writeFileSync(join(out, 'results.json'), JSON.stringify(results, null, 2) + '\n');
const cols = ['id', 'published', 'product', 'group', 'score', 'band', 'family', 'cwes', 'promptInDescription', 'kevAdded'];
const cell = (v) => { const s = v === null ? '' : String(v); return /[",\n]/.test(s) ? `"${s.replace(/"/g, '""')}"` : s; };
writeFileSync(join(out, 'ai-tooling-cves.csv'), [cols.join(','), ...rows.map((r) => cols.map((k) => cell(r[k])).join(','))].join('\n') + '\n');
console.log(`Done: ${rows.length} AI tooling CVEs out of ${allCves}.`);