sandbox native filesystem tools to prevent /proc/self/environ exfiltration (#509)

* sandbox native filesystem tools to prevent /proc/self/environ exfiltration the agent's native Read/Grep/Edit tools can bypass the shell sandbox by reading /proc/self/environ directly. this adds agent-native filesystem restrictions using the highest-precedence, non-overridable config for each CLI: OpenCode: OPENCODE_PERMISSION env var with external_directory deny-all + /tmp allow, plus deletion of untrusted .opencode/plugins/ and .opencode/tools/ before launch. Claude Code: managed-settings.json at /etc/claude-code/ with denyRead, permissions.deny, allowManagedPermissionRulesOnly, allowManagedHooksOnly. also --setting-sources user and --disallowedTools path patterns as belt-and-suspenders. Made-with: Cursor * add Glob to Claude Code /proc and /sys deny lists closes gap identified in review — Glob can enumerate /proc entries. added to both managed-settings.json permissions.deny and --disallowedTools. Made-with: Cursor * run token-exfil test with both agents, hint at native /proc reads changed tag from "agnostic" (opentoad-only) to "security" so the test runs with both opentoad and claude. updated prompt to explicitly instruct the agent to try reading /proc/self/environ via native Read tool. added API keys to action-agnostic CI job for claude support. Made-with: Cursor * move token-exfil to crossagent matrix, remove redundant permissions.deny - moved token-exfil from agnostic/ to crossagent/ so it runs via the agent matrix (claude + opentoad in parallel) instead of sequentially - removed permissions.deny per-tool rules from managed-settings.json; sandbox.filesystem.denyRead is the single enforcement mechanism - reverted action-agnostic env vars to minimal set - updated wiki to match Made-with: Cursor * document post-spawn API key deletion analysis in security wiki evaluated whether API key env vars can be deleted from agent processes after spawn. OpenCode snapshots env at startup (safe to delete), but Claude Code re-reads process.env per request (not viable). documented as further exploration item with per-agent breakdown and caveats. Made-with: Cursor * fix stale tokenExfil path references in wiki docs moved from test/agnostic/ to test/crossagent/ in directory tree and adversarial test example. Made-with: Cursor * revert accidental prisma.config.ts changes Made-with: Cursor * hardcode PULLFROG_MODEL per agent in test runner to avoid DB model mismatch when PULLFROG_AGENT forces a specific agent, the DB-configured model may belong to a different provider (e.g. openai model with claude agent). PULLFROG_MODEL short-circuits the DB slug resolution entirely. Made-with: Cursor
2026-04-02 22:31:41 +00:00
parent 36cc5cde14
commit 8cd36d221a
6 changed files with 149 additions and 20 deletions
@@ -3,6 +3,7 @@
 *
 * mirrors the opentoad harness's security model:
 * - native Bash blocked via --disallowedTools (agent cannot shell out)
+ * - managed-settings.json: filesystem sandbox — deny /proc, /sys reads
 * - MCP ShellTool provides restricted shell (filtered env, no secrets)
 * - MCP server injected via --mcp-config (not replacing project config)
 * - ASKPASS handles git auth separately (token never in subprocess env)
@@ -10,6 +11,7 @@
 * the agent process itself gets full env (needs LLM API keys, PATH, etc.).
 * security is enforced at the tool layer, not the process layer.
 */
+import { execFileSync } from "node:child_process";
 import { mkdirSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { performance } from "node:perf_hooks";
@@ -473,6 +475,57 @@ async function runClaude(params: RunParams): Promise<AgentResult> {
  }
 }

+// ── managed settings ────────────────────────────────────────────────────────────
+
+const MANAGED_SETTINGS_DIR = "/etc/claude-code";
+const MANAGED_SETTINGS_PATH = `${MANAGED_SETTINGS_DIR}/managed-settings.json`;
+
+// managed-settings.json has absolute highest precedence in Claude Code's config hierarchy.
+// it cannot be overridden by user, project, or local settings — safe against malicious PRs.
+//
+// permissions.deny blocks native tools (Read, Grep, Edit, Glob) from accessing /proc and /sys.
+// sandbox.filesystem.denyRead blocks the Bash tool sandbox from reading those paths.
+// allowManagedPermissionRulesOnly prevents malicious PRs from adding allow rules that override
+// our deny rules — safe in CI because --dangerously-skip-permissions makes allow/ask irrelevant.
+// allowManagedHooksOnly prevents malicious project hooks from bypassing deny rules.
+const managedSettings = {
+  allowManagedPermissionRulesOnly: true,
+  allowManagedHooksOnly: true,
+  permissions: {
+    deny: [
+      "Read(//proc/**)",
+      "Read(//sys/**)",
+      "Grep(//proc/**)",
+      "Grep(//sys/**)",
+      "Edit(//proc/**)",
+      "Edit(//sys/**)",
+      "Glob(//proc/**)",
+      "Glob(//sys/**)",
+    ],
+  },
+  sandbox: {
+    filesystem: {
+      denyRead: ["/proc", "/sys"],
+    },
+  },
+};
+
+function installManagedSettings(): void {
+  if (process.env.CI !== "true") return;
+
+  const content = JSON.stringify(managedSettings, null, 2);
+  try {
+    execFileSync("sudo", ["mkdir", "-p", MANAGED_SETTINGS_DIR]);
+    execFileSync("sudo", ["tee", MANAGED_SETTINGS_PATH], {
+      input: content,
+      stdio: ["pipe", "ignore", "pipe"],
+    });
+    log.debug(`» wrote managed settings to ${MANAGED_SETTINGS_PATH}`);
+  } catch (err) {
+    log.warning(`» failed to install managed settings: ${err}`);
+  }
+}
+
 // ── agent ───────────────────────────────────────────────────────────────────────

 export const claude = agent({
@@ -502,6 +555,8 @@ export const claude = agent({
    const mcpConfigPath = writeMcpConfig(ctx);
    const effort = resolveEffort(model);

+    installManagedSettings();
+
    const args = [
      cliPath,
      "-p",
@@ -525,7 +580,7 @@ export const claude = agent({
    }

    // agent process gets full env — needs LLM API keys, PATH, locale, etc.
-    // security is enforced via --disallowedTools (Bash + Bash subagent) and MCP tool filtering.
+    // security is enforced via managed-settings.json, --disallowedTools (Bash), and MCP tool filtering.
    const env: Record<string, string | undefined> = {
      ...process.env,
      ...homeEnv,
@@ -3,6 +3,8 @@
 *
 * transparently wraps OpenCode with a security layer:
 * - bash: "deny" via OPENCODE_CONFIG_CONTENT (agent cannot shell out)
+ * - OPENCODE_PERMISSION: filesystem sandbox — deny all external paths except /tmp
+ * - untrusted .opencode/plugins/ and .opencode/tools/ deleted before launch
 * - MCP ShellTool provides restricted shell (filtered env, no secrets)
 * - MCP server injected alongside project config (not replacing)
 * - ASKPASS handles git auth separately (token never in subprocess env)
@@ -11,7 +13,7 @@
 * security is enforced at the tool layer, not the process layer.
 */
 import { execFileSync } from "node:child_process";
-import { mkdirSync } from "node:fs";
+import { mkdirSync, rmSync } from "node:fs";
 import { join } from "node:path";
 import { performance } from "node:perf_hooks";
 import { ghPullfrogMcpName } from "../external.ts";
@@ -625,14 +627,24 @@ export const opentoad = agent({
      agent: "opencode",
    });

+    // remove untrusted project plugins/tools before agent starts (no env var to disable loading)
+    rmSync(join(process.cwd(), ".opencode", "plugins"), { recursive: true, force: true });
+    rmSync(join(process.cwd(), ".opencode", "tools"), { recursive: true, force: true });
+
    const args = ["run", ctx.instructions.full, "--format", "json", "--print-logs"];

-    // agent process gets full env — needs LLM API keys, PATH, locale, etc.
-    // security is enforced via OPENCODE_CONFIG_CONTENT (bash: deny) and MCP tool filtering.
+    // OPENCODE_PERMISSION has absolute highest precedence (merged after managed/MDM configs).
+    // external_directory gates ALL native filesystem tools (Read, Write, Edit, Glob, Grep, etc.)
+    // for paths outside the project root. last-match-wins: deny everything, then allow /tmp.
+    const permissionOverride = JSON.stringify({
+      external_directory: { "*": "deny", "/tmp/*": "allow" },
+    });
+
    const env: Record<string, string | undefined> = {
      ...process.env,
      ...homeEnv,
      OPENCODE_CONFIG_CONTENT: buildSecurityConfig(ctx, model),
+      OPENCODE_PERMISSION: permissionOverride,
      GOOGLE_GENERATIVE_AI_API_KEY:
        process.env.GOOGLE_GENERATIVE_AI_API_KEY || process.env.GEMINI_API_KEY,
    };