mirror of
https://github.com/openclaw/openclaw.git
synced 2026-04-03 03:03:24 -04:00
Agents/Tools: preflight exec script files for shell var injection (#18457)
* fix(agents): don't force store=true for codex responses * test: stabilize respawn + subagent usage assertions * Agents/Tools: preflight exec to detect shell variable injection in scripts * Changelog: fix merge marker formatting
This commit is contained in:
@@ -14,6 +14,7 @@ Docs: https://docs.openclaw.ai
|
||||
|
||||
### Fixes
|
||||
|
||||
- Agents/Tools/exec: add a preflight guard that detects likely shell env var injection (e.g. `$DM_JSON`, `$TMPDIR`) in Python/Node scripts before execution, preventing recurring cron failures and wasted tokens when models emit mixed shell+language source. (#12836)
|
||||
- Security/Sessions: create new session transcript JSONL files with user-only (`0o600`) permissions and extend `openclaw security audit --fix` to remediate existing transcript file permissions.
|
||||
- Infra/Fetch: ensure foreign abort-signal listener cleanup never masks original fetch successes/failures, while still preventing detached-finally unhandled rejection noise in `wrapFetchWithAbortSignal`. Thanks @Jackten.
|
||||
- Gateway/Config: prevent `config.patch` object-array merges from falling back to full-array replacement when some patch entries lack `id`, so partial `agents.list` updates no longer drop unrelated agents. (#17989) Thanks @stakeswky.
|
||||
|
||||
65
src/agents/bash-tools.exec.script-preflight.test.ts
Normal file
65
src/agents/bash-tools.exec.script-preflight.test.ts
Normal file
@@ -0,0 +1,65 @@
|
||||
import fs from "node:fs/promises";
|
||||
import os from "node:os";
|
||||
import path from "node:path";
|
||||
import { describe, expect, it } from "vitest";
|
||||
|
||||
const isWin = process.platform === "win32";
|
||||
|
||||
describe("exec script preflight", () => {
|
||||
it("blocks shell env var injection tokens in python scripts before execution", async () => {
|
||||
if (isWin) {
|
||||
return;
|
||||
}
|
||||
|
||||
const tmp = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-exec-preflight-"));
|
||||
const pyPath = path.join(tmp, "bad.py");
|
||||
|
||||
await fs.writeFile(
|
||||
pyPath,
|
||||
[
|
||||
"import json",
|
||||
"# model accidentally wrote shell syntax:",
|
||||
"payload = $DM_JSON",
|
||||
"print(payload)",
|
||||
].join("\n"),
|
||||
"utf-8",
|
||||
);
|
||||
|
||||
const { createExecTool } = await import("./bash-tools.exec.js");
|
||||
const tool = createExecTool({ host: "gateway", security: "full", ask: "off" });
|
||||
|
||||
await expect(
|
||||
tool.execute("call1", {
|
||||
command: "python bad.py",
|
||||
workdir: tmp,
|
||||
}),
|
||||
).rejects.toThrow(/exec preflight: detected likely shell variable injection \(\$DM_JSON\)/);
|
||||
});
|
||||
|
||||
it("blocks obvious shell-as-js output before node execution", async () => {
|
||||
if (isWin) {
|
||||
return;
|
||||
}
|
||||
|
||||
const tmp = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-exec-preflight-"));
|
||||
const jsPath = path.join(tmp, "bad.js");
|
||||
|
||||
await fs.writeFile(
|
||||
jsPath,
|
||||
['NODE "$TMPDIR/hot.json"', "console.log('hi')"].join("\n"),
|
||||
"utf-8",
|
||||
);
|
||||
|
||||
const { createExecTool } = await import("./bash-tools.exec.js");
|
||||
const tool = createExecTool({ host: "gateway", security: "full", ask: "off" });
|
||||
|
||||
await expect(
|
||||
tool.execute("call1", {
|
||||
command: "node bad.js",
|
||||
workdir: tmp,
|
||||
}),
|
||||
).rejects.toThrow(
|
||||
/exec preflight: (detected likely shell variable injection|JS file starts with shell syntax)/,
|
||||
);
|
||||
});
|
||||
});
|
||||
@@ -1,5 +1,7 @@
|
||||
import type { AgentTool, AgentToolResult } from "@mariozechner/pi-agent-core";
|
||||
import crypto from "node:crypto";
|
||||
import fs from "node:fs/promises";
|
||||
import path from "node:path";
|
||||
import type { BashSandboxConfig } from "./bash-tools.shared.js";
|
||||
import {
|
||||
type ExecAsk,
|
||||
@@ -118,6 +120,97 @@ export type ExecToolDetails =
|
||||
nodeId?: string;
|
||||
};
|
||||
|
||||
function extractScriptTargetFromCommand(
|
||||
command: string,
|
||||
): { kind: "python"; relOrAbsPath: string } | { kind: "node"; relOrAbsPath: string } | null {
|
||||
const raw = command.trim();
|
||||
if (!raw) {
|
||||
return null;
|
||||
}
|
||||
|
||||
// Intentionally simple parsing: we only support common forms like
|
||||
// python file.py
|
||||
// python3 -u file.py
|
||||
// node --experimental-something file.js
|
||||
// If the command is more complex (pipes, heredocs, quoted paths with spaces), skip preflight.
|
||||
const pythonMatch = raw.match(/^\s*(python3?|python)\s+(?:-[^\s]+\s+)*([^\s]+\.py)\b/i);
|
||||
if (pythonMatch?.[2]) {
|
||||
return { kind: "python", relOrAbsPath: pythonMatch[2] };
|
||||
}
|
||||
const nodeMatch = raw.match(/^\s*(node)\s+(?:--[^\s]+\s+)*([^\s]+\.js)\b/i);
|
||||
if (nodeMatch?.[2]) {
|
||||
return { kind: "node", relOrAbsPath: nodeMatch[2] };
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
|
||||
async function validateScriptFileForShellBleed(params: {
|
||||
command: string;
|
||||
workdir: string;
|
||||
}): Promise<void> {
|
||||
const target = extractScriptTargetFromCommand(params.command);
|
||||
if (!target) {
|
||||
return;
|
||||
}
|
||||
|
||||
const absPath = path.isAbsolute(target.relOrAbsPath)
|
||||
? path.resolve(target.relOrAbsPath)
|
||||
: path.resolve(params.workdir, target.relOrAbsPath);
|
||||
|
||||
// Best-effort: only validate if file exists and is reasonably small.
|
||||
let stat: { isFile(): boolean; size: number };
|
||||
try {
|
||||
stat = await fs.stat(absPath);
|
||||
} catch {
|
||||
return;
|
||||
}
|
||||
if (!stat.isFile()) {
|
||||
return;
|
||||
}
|
||||
if (stat.size > 512 * 1024) {
|
||||
return;
|
||||
}
|
||||
|
||||
const content = await fs.readFile(absPath, "utf-8");
|
||||
|
||||
// Common failure mode: shell env var syntax leaking into Python/JS.
|
||||
// We deliberately match all-caps/underscore vars to avoid false positives with `$` as a JS identifier.
|
||||
const envVarRegex = /\$[A-Z_][A-Z0-9_]{1,}/g;
|
||||
const first = envVarRegex.exec(content);
|
||||
if (first) {
|
||||
const idx = first.index;
|
||||
const before = content.slice(0, idx);
|
||||
const line = before.split("\n").length;
|
||||
const token = first[0];
|
||||
throw new Error(
|
||||
[
|
||||
`exec preflight: detected likely shell variable injection (${token}) in ${target.kind} script: ${path.basename(
|
||||
absPath,
|
||||
)}:${line}.`,
|
||||
target.kind === "python"
|
||||
? `In Python, use os.environ.get(${JSON.stringify(token.slice(1))}) instead of raw ${token}.`
|
||||
: `In Node.js, use process.env[${JSON.stringify(token.slice(1))}] instead of raw ${token}.`,
|
||||
"(If this is inside a string literal on purpose, escape it or restructure the code.)",
|
||||
].join("\n"),
|
||||
);
|
||||
}
|
||||
|
||||
// Another recurring pattern from the issue: shell commands accidentally emitted as JS.
|
||||
if (target.kind === "node") {
|
||||
const firstNonEmpty = content
|
||||
.split(/\r?\n/)
|
||||
.map((l) => l.trim())
|
||||
.find((l) => l.length > 0);
|
||||
if (firstNonEmpty && /^NODE\b/.test(firstNonEmpty)) {
|
||||
throw new Error(
|
||||
`exec preflight: JS file starts with shell syntax (${firstNonEmpty}). ` +
|
||||
`This looks like a shell command, not JavaScript.`,
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
export function createExecTool(
|
||||
defaults?: ExecToolDefaults,
|
||||
// oxlint-disable-next-line typescript/no-explicit-any
|
||||
@@ -874,6 +967,11 @@ export function createExecTool(
|
||||
typeof params.timeout === "number" ? params.timeout : defaultTimeoutSec;
|
||||
const getWarningText = () => (warnings.length ? `${warnings.join("\n")}\n\n` : "");
|
||||
const usePty = params.pty === true && !sandbox;
|
||||
|
||||
// Preflight: catch a common model failure mode (shell syntax leaking into Python/JS sources)
|
||||
// before we execute and burn tokens in cron loops.
|
||||
await validateScriptFileForShellBleed({ command: params.command, workdir });
|
||||
|
||||
const run = await runExecProcess({
|
||||
command: params.command,
|
||||
execCommand: execCommandOverride,
|
||||
|
||||
@@ -154,4 +154,27 @@ describe("applyExtraParamsToAgent", () => {
|
||||
});
|
||||
expect(payload.store).toBe(false);
|
||||
});
|
||||
|
||||
it("does not force store=true for Codex responses (Codex requires store=false)", () => {
|
||||
const payload = { store: false };
|
||||
const baseStreamFn: StreamFn = (_model, _context, options) => {
|
||||
options?.onPayload?.(payload);
|
||||
return new AssistantMessageEventStream();
|
||||
};
|
||||
const agent = { streamFn: baseStreamFn };
|
||||
|
||||
applyExtraParamsToAgent(agent, undefined, "openai-codex", "codex-mini-latest");
|
||||
|
||||
const model = {
|
||||
api: "openai-codex-responses",
|
||||
provider: "openai-codex",
|
||||
id: "codex-mini-latest",
|
||||
baseUrl: "https://chatgpt.com/backend-api/codex/responses",
|
||||
} as Model<"openai-codex-responses">;
|
||||
const context: Context = { messages: [] };
|
||||
|
||||
void agent.streamFn?.(model, context, {});
|
||||
|
||||
expect(payload.store).toBe(false);
|
||||
});
|
||||
});
|
||||
|
||||
Reference in New Issue
Block a user