0sec-cli 0.15.0 → 0.18.0

package/chunks/dist-FMGSR3DW.js

@@ -0,0 +1,159 @@
+#!/usr/bin/env node
+import { createRequire as __0secCreateRequire } from "node:module";
+const require = __0secCreateRequire(import.meta.url);
+import {
+  AGENT_ACTION_ASSURANCE_SCHEMA_VERSION,
+  AgentActionEvidenceManifestSchema,
+  AgentActionOutcomeSchema,
+  AgentActionScenarioSchema,
+  AgentActionTargetMetadataSchema,
+  AgentInjectionVectorSchema,
+  AssertionKindSchema,
+  DEFAULT_ALLOW_MODEL_SELF_EXTENSION,
+  DEFAULT_AUTONOMY_MODE,
+  DEFAULT_MAX_CONCURRENCY,
+  DEFAULT_MODEL,
+  DEFAULT_SEVERITY_FLOOR,
+  DEFAULT_TIMEOUT_MS,
+  DEPTH_CONFIG,
+  DESKTOP_CONSOLE_SCHEMA_VERSION,
+  EngineMetadataSchema,
+  EvidenceArtifactSchema,
+  HALT_STATE,
+  MANUAL_PRICING,
+  MODEL_PRICING,
+  PRESENTATION_PROTOCOL,
+  PRICING_SNAPSHOT_DATE,
+  RunnerKindSchema,
+  SEVERITY_ORDER,
+  SEVERITY_RANK,
+  VERSION,
+  VerificationAssertionSchema,
+  VerificationCommandSchema,
+  VerificationModeSchema,
+  VerificationResultSchema,
+  VerificationStatusSchema,
+  compareRoles,
+  conditionCmpOpSchema,
+  conditionCmpSchema,
+  conditionExprSchema,
+  conditionLhsSchema,
+  conditionRhsSchema,
+  createAuditReportDocument,
+  createPresentationEvent,
+  createReviewReportDocument,
+  createScanReportDocument,
+  createTranscriptDocument,
+  defaultRoleFor,
+  estimateCost,
+  findingSchema,
+  formatZodError,
+  getRates,
+  hasMultipleIdentities,
+  homeStateDir,
+  isPresentationEvent,
+  meetsSeverityFloor,
+  modelProvider,
+  normalizeResearchNovelty,
+  parseAgentActionEvidenceManifest,
+  pocStepArraySchema,
+  pocStepSchema,
+  priceRun,
+  reportSummarySchema,
+  researchDisclosureReady,
+  researchGradeAtLeast,
+  researchLpeDisclosureReady,
+  researchPlatformLpeDisclosureReady,
+  researchWindowsLpeDisclosureReady,
+  researchWindowsTokenTransitionProven,
+  researchZeroCapProven,
+  resolveIdentities,
+  runStateDir,
+  severityRank,
+  specialistRefSchema,
+  splitCost,
+  stateDefSchema,
+  transitionRefSchema,
+  transitionRuleSchema,
+  transitionTargetSchema,
+  workflowDefinitionSchema
+} from "./chunk-6L26OBWI.js";
+import "./chunk-YLMN3N25.js";
+import "./chunk-IR537GON.js";
+export {
+  AGENT_ACTION_ASSURANCE_SCHEMA_VERSION,
+  AgentActionEvidenceManifestSchema,
+  AgentActionOutcomeSchema,
+  AgentActionScenarioSchema,
+  AgentActionTargetMetadataSchema,
+  AgentInjectionVectorSchema,
+  AssertionKindSchema,
+  DEFAULT_ALLOW_MODEL_SELF_EXTENSION,
+  DEFAULT_AUTONOMY_MODE,
+  DEFAULT_MAX_CONCURRENCY,
+  DEFAULT_MODEL,
+  DEFAULT_SEVERITY_FLOOR,
+  DEFAULT_TIMEOUT_MS,
+  DEPTH_CONFIG,
+  DESKTOP_CONSOLE_SCHEMA_VERSION,
+  EngineMetadataSchema,
+  EvidenceArtifactSchema,
+  HALT_STATE,
+  MANUAL_PRICING,
+  MODEL_PRICING,
+  PRESENTATION_PROTOCOL,
+  PRICING_SNAPSHOT_DATE,
+  RunnerKindSchema,
+  SEVERITY_ORDER,
+  SEVERITY_RANK,
+  VERSION,
+  VerificationAssertionSchema,
+  VerificationCommandSchema,
+  VerificationModeSchema,
+  VerificationResultSchema,
+  VerificationStatusSchema,
+  compareRoles,
+  conditionCmpOpSchema,
+  conditionCmpSchema,
+  conditionExprSchema,
+  conditionLhsSchema,
+  conditionRhsSchema,
+  createAuditReportDocument,
+  createPresentationEvent,
+  createReviewReportDocument,
+  createScanReportDocument,
+  createTranscriptDocument,
+  defaultRoleFor,
+  estimateCost,
+  findingSchema,
+  formatZodError,
+  getRates,
+  hasMultipleIdentities,
+  homeStateDir,
+  isPresentationEvent,
+  meetsSeverityFloor,
+  modelProvider,
+  normalizeResearchNovelty,
+  parseAgentActionEvidenceManifest,
+  pocStepArraySchema,
+  pocStepSchema,
+  priceRun,
+  reportSummarySchema,
+  researchDisclosureReady,
+  researchGradeAtLeast,
+  researchLpeDisclosureReady,
+  researchPlatformLpeDisclosureReady,
+  researchWindowsLpeDisclosureReady,
+  researchWindowsTokenTransitionProven,
+  researchZeroCapProven,
+  resolveIdentities,
+  runStateDir,
+  severityRank,
+  specialistRefSchema,
+  splitCost,
+  stateDefSchema,
+  transitionRefSchema,
+  transitionRuleSchema,
+  transitionTargetSchema,
+  workflowDefinitionSchema
+};

package/chunks/eval-runner-7ZF472KV.js

@@ -0,0 +1,27 @@
+#!/usr/bin/env node
+import { createRequire as __0secCreateRequire } from "node:module";
+const require = __0secCreateRequire(import.meta.url);
+import {
+  getEvalCategories,
+  runEval
+} from "./chunk-AHTBZC3E.js";
+import "./chunk-BZNEY2ZS.js";
+import "./chunk-YYRBQVFE.js";
+import "./chunk-3MOLBTLS.js";
+import "./chunk-P6WKNFWX.js";
+import "./chunk-UCYRA73C.js";
+import "./chunk-LP3HHYQU.js";
+import "./chunk-KLTNTE2Z.js";
+import "./chunk-RRMJC3ZE.js";
+import "./chunk-H2FFLZNK.js";
+import "./chunk-WEBMRWEG.js";
+import "./chunk-57ZENEX2.js";
+import "./chunk-UR4ELX2Y.js";
+import "./chunk-ADOLI6DT.js";
+import "./chunk-6L26OBWI.js";
+import "./chunk-YLMN3N25.js";
+import "./chunk-IR537GON.js";
+export {
+  getEvalCategories,
+  runEval
+};

package/chunks/example-manifest.json

@@ -0,0 +1,95 @@
+{
+  "id": "bench-example-references-only",
+  "version": 1,
+  "corpusRoot": "/opt/0sec-bench-corpus",
+  "_comment": "REFERENCES ONLY. This sample documents the manifest shape. It contains NO vulnerable app source and NO exploits. The real uncontaminated/post-cutoff corpus (images, compose dirs, kernel reproducers) lives OUTSIDE this repo per disclosure embargo discipline and is resolved at run time via image refs / corpusRoot-relative paths / reproducerRef. Markers below are placeholder tokens that, in a real corpus, are planted per-target at build time (like XBOW's random FLAG).",
+  "cases": [
+    {
+      "id": "web-file-read-001",
+      "name": "Path traversal → /etc/secret read",
+      "target": { "kind": "web", "image": "ghcr.io/0sec-labs/bench-web-file-read-001@sha256:PINNED", "port": 8080 },
+      "objective": { "type": "file-read", "marker": "BENCH_MARKER_FILEREAD_001", "path": "/etc/0sec-bench/secret" },
+      "ci": true,
+      "tags": ["web", "lfi"]
+    },
+    {
+      "id": "web-db-access-002",
+      "name": "SQLi → exfiltrate seeded row",
+      "target": { "kind": "web", "composeDir": "web/db-access-002", "port": 3000 },
+      "objective": { "type": "db-access", "marker": "BENCH_MARKER_DBROW_002" },
+      "ci": true,
+      "tags": ["web", "sqli"]
+    },
+    {
+      "id": "web-admin-login-003",
+      "name": "Auth bypass → admin panel",
+      "target": { "kind": "web", "image": "ghcr.io/0sec-labs/bench-web-admin-003@sha256:PINNED", "port": 80 },
+      "objective": { "type": "admin-login", "marker": "BENCH_MARKER_ADMINFLAG_003" },
+      "tags": ["web", "authz"]
+    },
+    {
+      "id": "web-ssrf-callback-004",
+      "name": "SSRF → attacker callback",
+      "target": { "kind": "web", "image": "ghcr.io/0sec-labs/bench-web-ssrf-004@sha256:PINNED", "port": 8000 },
+      "objective": { "type": "ssrf-callback", "token": "BENCH_OAST_TOKEN_004" },
+      "tags": ["web", "ssrf"]
+    },
+    {
+      "id": "web-file-write-005",
+      "name": "Upload → arbitrary write",
+      "target": { "kind": "web", "composeDir": "web/file-write-005", "port": 5000 },
+      "objective": { "type": "file-write", "marker": "BENCH_MARKER_FILEWRITE_005", "path": "/var/www/uploads/pwn" },
+      "tags": ["web", "upload"]
+    },
+
+    {
+      "id": "web-negative-hardened-101",
+      "name": "Hardened search endpoint (parameterized queries)",
+      "target": { "kind": "web", "image": "ghcr.io/0sec-labs/bench-web-negative-101@sha256:PINNED", "port": 8080 },
+      "objective": { "type": "db-access", "marker": "BENCH_MARKER_NEVER_PRESENT_101" },
+      "knownNegative": true,
+      "ci": true,
+      "tags": ["web", "known-negative"]
+    },
+    {
+      "id": "web-negative-allowlist-102",
+      "name": "URL fetcher with strict allowlist (no SSRF)",
+      "target": { "kind": "web", "image": "ghcr.io/0sec-labs/bench-web-negative-102@sha256:PINNED", "port": 8000 },
+      "objective": { "type": "ssrf-callback", "token": "BENCH_OAST_TOKEN_NEVER_102" },
+      "knownNegative": true,
+      "tags": ["web", "known-negative"]
+    },
+    {
+      "id": "web-negative-rbac-103",
+      "name": "Correct RBAC on admin routes (no bypass)",
+      "target": { "kind": "web", "image": "ghcr.io/0sec-labs/bench-web-negative-103@sha256:PINNED", "port": 80 },
+      "objective": { "type": "admin-login", "marker": "BENCH_MARKER_NEVER_104" },
+      "knownNegative": true,
+      "tags": ["web", "known-negative"]
+    },
+
+    {
+      "id": "kernel-uaf-201",
+      "name": "Driver UAF reproducer (KASAN slab-use-after-free)",
+      "target": { "kind": "kernel", "reproducerRef": "corpus://kernel/uaf-201", "ecosystem": "kernel-tree" },
+      "objective": { "type": "kasan-hit", "signature": "slab-use-after-free" },
+      "ci": true,
+      "tags": ["kernel", "uaf"]
+    },
+    {
+      "id": "kernel-oob-202",
+      "name": "Out-of-bounds write reproducer (KASAN)",
+      "target": { "kind": "kernel", "reproducerRef": "corpus://kernel/oob-202", "ecosystem": "kernel-tree" },
+      "objective": { "type": "kasan-hit", "signature": "out-of-bounds" },
+      "tags": ["kernel", "oob"]
+    },
+    {
+      "id": "kernel-negative-203",
+      "name": "Patched driver (no crash) — known-negative",
+      "target": { "kind": "kernel", "reproducerRef": "corpus://kernel/patched-203", "ecosystem": "kernel-tree" },
+      "objective": { "type": "kasan-hit" },
+      "knownNegative": true,
+      "tags": ["kernel", "known-negative"]
+    }
+  ]
+}

package/chunks/exploit-agent-45YQOHPJ.js

@@ -0,0 +1,98 @@
+#!/usr/bin/env node
+import { createRequire as __0secCreateRequire } from "node:module";
+const require = __0secCreateRequire(import.meta.url);
+import "./chunk-IR537GON.js";
+
+// packages/cli/src/commands/exploit-agent.ts
+import { execFileSync } from "node:child_process";
+import { writeFileSync } from "node:fs";
+import { resolve } from "node:path";
+function makeDockerExecutor(containerId, timeoutMs = 3e5) {
+  return (cmd) => execFileSync("docker", ["exec", "-w", "/workspace", containerId, "sh", "-c", cmd], {
+    encoding: "utf8",
+    timeout: timeoutMs,
+    maxBuffer: 64 * 1024 * 1024
+  });
+}
+function makeScriptExecutor(scriptPath, timeoutMs = 3e5) {
+  const abs = resolve(scriptPath);
+  return (cmd) => execFileSync(abs, [cmd], {
+    encoding: "utf8",
+    timeout: timeoutMs,
+    maxBuffer: 64 * 1024 * 1024
+  });
+}
+async function runExploitAgent(opts) {
+  const { runExploitScan } = await import("./dist-4YICS2J2.js");
+  const log = opts.log ?? ((m) => process.stderr.write(m + "\n"));
+  const result = await runExploitScan({
+    target: {
+      description: opts.taskDescription,
+      flagPath: opts.flagPath,
+      ...opts.flagPattern ? { flagPattern: opts.flagPattern } : {},
+      ...opts.taskId ? { taskId: opts.taskId } : {}
+    },
+    execCmd: opts.exec,
+    runtime: opts.runtime ?? "api",
+    ...opts.model ? { model: opts.model } : {},
+    ...opts.maxSteps !== void 0 ? { maxSteps: opts.maxSteps } : {},
+    log
+  });
+  const exitCode = result.captured ? 0 : result.builtExploit ? 1 : 2;
+  return {
+    exitCode,
+    result: {
+      mode: "agent",
+      ...opts.taskId ? { task_id: opts.taskId } : {},
+      captured: result.captured,
+      steps: result.steps,
+      built_exploit: result.builtExploit,
+      model: result.model,
+      flag_path: opts.flagPath,
+      // Surfaced deliberately: the loop warns here when a "capture" rests only on
+      // a non-empty flag file rather than a real oracle.
+      warnings: result.warnings
+    }
+  };
+}
+async function agentFromExploitOpts(opts) {
+  if (!opts.task) throw new Error("--agent requires --task <path to the vuln/task description>");
+  if (!opts.container && !opts.execScript) {
+    throw new Error(
+      "--agent requires a target executor: --container <id> (docker exec) or --exec-script <path>. Refusing to run exploit commands locally."
+    );
+  }
+  if (opts.container && opts.execScript) {
+    throw new Error("--agent takes --container OR --exec-script, not both");
+  }
+  const { readFileSync } = await import("node:fs");
+  const taskDescription = readFileSync(resolve(opts.task), "utf8");
+  const flagPath = opts.flagPath ?? "/tmp/flag";
+  let maxSteps;
+  if (opts.maxSteps !== void 0) {
+    maxSteps = parseInt(opts.maxSteps, 10);
+    if (!Number.isFinite(maxSteps) || maxSteps <= 0) {
+      throw new Error(`invalid --max-steps '${opts.maxSteps}' (expected positive integer)`);
+    }
+  }
+  const exec = opts.container ? makeDockerExecutor(opts.container) : makeScriptExecutor(opts.execScript);
+  const outcome = await runExploitAgent({
+    taskDescription,
+    flagPath,
+    ...opts.flagPattern ? { flagPattern: opts.flagPattern } : {},
+    exec,
+    ...opts.model ? { model: opts.model } : {},
+    ...maxSteps !== void 0 ? { maxSteps } : {},
+    ...opts.runtime ? { runtime: opts.runtime } : {}
+  });
+  const json = JSON.stringify(outcome.result, null, 2);
+  if (opts.output) writeFileSync(resolve(opts.output), json + "\n", "utf8");
+  else process.stdout.write(json + "\n");
+  return outcome.exitCode;
+}
+export {
+  agentFromExploitOpts,
+  makeDockerExecutor,
+  makeScriptExecutor,
+  runExploitAgent
+};

package/chunks/exploit-autoclimb-UZGEMGYT.js

@@ -0,0 +1,81 @@
+#!/usr/bin/env node
+import { createRequire as __0secCreateRequire } from "node:module";
+const require = __0secCreateRequire(import.meta.url);
+import "./chunk-IR537GON.js";
+
+// packages/cli/src/commands/exploit-autoclimb.ts
+import { spawn } from "node:child_process";
+import { mkdtempSync, writeFileSync, readFileSync, existsSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+function makeShellBootRunner(opts) {
+  const log = opts.logger ?? (() => {
+  });
+  if (!existsSync(opts.bootScript)) {
+    throw new Error(`autoclimb boot script not found: ${opts.bootScript}`);
+  }
+  return async (report) => {
+    const dir = mkdtempSync(join(tmpdir(), "autoclimb-"));
+    const cfile = join(dir, "poc.c");
+    writeFileSync(cfile, report.reproducer ?? "", "utf8");
+    const serialFile = `${cfile}.serial`;
+    log(`[autoclimb-runner] booting ${cfile} via ${opts.bootScript}`);
+    const output = await new Promise((resolveP) => {
+      const child = spawn(opts.bootScript, [cfile], {
+        env: { ...process.env, ...opts.env },
+        timeout: opts.timeoutMs ?? 9e5
+      });
+      let out = "";
+      child.stdout.on("data", (d) => out += d.toString());
+      child.stderr.on("data", (d) => out += d.toString());
+      child.on("close", () => resolveP(out));
+      child.on("error", (e) => resolveP(`boot-script error: ${String(e)}`));
+    });
+    const dmesg = existsSync(serialFile) ? readFileSync(serialFile, "utf8") : "";
+    const compiled = /COMPILED=1/.test(output);
+    const executed = /EXECUTED=1/.test(output);
+    return { compiled, executed, output, dmesg, exitCode: 0, timedOut: false };
+  };
+}
+async function runAutoclimbCli(opts) {
+  const log = opts.logger ?? ((l) => process.stderr.write(l + "\n"));
+  const { autonomousClimb } = await import("./dist-4YICS2J2.js");
+  const bug = JSON.parse(readFileSync(opts.bugSpecPath, "utf8"));
+  const runner = makeShellBootRunner({
+    bootScript: opts.bootScript,
+    ...opts.bootEnv ? { env: opts.bootEnv } : {},
+    logger: log
+  });
+  const result = await autonomousClimb({
+    bug,
+    runner,
+    maxRounds: opts.maxRounds,
+    plateauRounds: opts.plateauRounds,
+    ...opts.modelId ? { modelId: opts.modelId } : {},
+    logger: log
+  });
+  return result;
+}
+async function autoclimbFromExploitOpts(opts) {
+  const bugSpecPath = opts.bugSpec;
+  const bootScript = opts.bootScript ?? process.env["0SEC_AUTOCLIMB_BOOT_SCRIPT"];
+  if (!bugSpecPath) throw new Error("--autoclimb requires --bug-spec <path>");
+  if (!bootScript) throw new Error("--autoclimb requires --boot-script <path> or 0SEC_AUTOCLIMB_BOOT_SCRIPT");
+  const maxRounds = opts.loopBoots ? parseInt(opts.loopBoots, 10) : 6;
+  const result = await runAutoclimbCli({
+    bugSpecPath,
+    bootScript,
+    maxRounds: Number.isFinite(maxRounds) && maxRounds > 0 ? maxRounds : 6,
+    plateauRounds: 3,
+    ...opts.model ? { modelId: opts.model } : {}
+  });
+  const json = JSON.stringify(result, null, 2);
+  if (opts.output) writeFileSync(opts.output, json + "\n", "utf8");
+  else process.stdout.write(json + "\n");
+  return result.reachedCeiling ? 0 : result.highestRung === "none" ? 2 : 1;
+}
+export {
+  autoclimbFromExploitOpts,
+  makeShellBootRunner,
+  runAutoclimbCli
+};