Browse / Direct Prompt Injection / CVE-2026-57129
PraisonAI: Arbitrary File Read via `@file:` Mention Path Traversal
- Class
- Direct Prompt Injection (filed by a model that read the advisory)
- Technique
- Path traversal and file access
- Also involves
- Tool and Agent Misuse
- Affected
- praisonaiagents (PyPI)
- Gets past
- Path restriction
- CVSS
- CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N
- EPSS
- 0.5% chance of exploitation in 30 days
How it works, in short
Prompt-level @file mentions fall back to unchecked absolute paths, letting an attacker read any file the agent process can access. (Our summary, written by a model. The source text is below.)
Details
Summary
The MentionsParser in src/praisonai-agents/praisonaiagents/tools/mentions.py processes @file: mentions in agent prompts by reading arbitrary files from the filesystem. When a file path is not found relative to the workspace, the parser falls back to using the path as an absolute path without any validation or boundary check. This allows an attacker who can influence agent prompts (via chat messages, Telegram/Discord/Slack bot inputs, or YAML workflow configs) to read any file on the filesystem accessible to the process user.
Details
Vulnerable code (lines 165–178):
def _process_file_mention(self, file_path: str) -> Optional[str]:
"""Process @file:path mention."""
try:
# Resolve path relative to workspace
full_path = self.workspace_path / file_path
if not full_path.exists():
# Try as absolute path
full_path = Path(file_path)
if not full_path.exists():
self._log(f"File not found: {file_path}", logging.WARNING)
return f"# File: {file_path}\n[File not found]"
content = full_path.read_text(encoding="utf-8")
The vulnerability is in the fallback at line 171–172: When the file is not found relative to workspace_path, the code constructs full_path = Path(file_path), which accepts any absolute or relative path without validation. There is no:
- .. path traversal check
- Workspace boundary validation
- Symlink resolution against workspace
- Protected path guard
The file_path parameter originates from parsing @file: mentions in user/LLM prompts. The MentionsParser is used across the framework to process mentions in agent instructions and user messages.
Contrast with skill_tools.py read_skill_file (lines 140–193), which properly validates:
# skill_tools.py line 179 — proper validation
if os.path.commonpath([full_path, skill_path]) != skill_path:
return f"Error: Path traversal detected - {file_path} is outside skill directory"
PoC
Setup: Clean checkout at commit d5f1114a.
Positive trigger — arbitrary file read via @file: mention:
import sys
sys.path.insert(0, 'src/praisonai-agents')
from praisonaiagents.tools.mentions import MentionsParser
parser = MentionsParser()
# Test 1: Absolute path read (bypasses workspace resolution)
result = parser._process_file_mention('/etc/hostname')
print(f'Absolute path read: {result[:80]}...')
# Test 2: Relative path with traversal
result = parser._process_file_mention('../../../etc/hostname')
print(f'Traversal read: {result[:80]}...')
Expected output:
Absolute path read: # File: /etc/hostname
linux
<hostname>
Traversal read: # File: ../../../etc/hostname
linux
<hostname>
Negative control — non-existent file:
result = parser._process_file_mention('/nonexistent/secret.txt')
# Returns: "# File: /nonexistent/secret.txt\n[File not found]"
Cleanup: No persistence or side effects — read-only operation.
Impact
An attacker who can inject @file: mentions into agent prompts (via chat messages in Telegram/Discord/Slack bots, user input in web UI, or YAML workflow configurations) can read any file accessible to the process user, including:
- Secrets and credentials:
.envfiles,~/.aws/credentials,~/.ssh/id_rsa, API keys - Configuration files: Database passwords, JWT secrets, OAuth tokens
- Source code: Application internals, database schemas
- System files:
/etc/passwd,/etc/shadow(if process has read access)
This is particularly dangerous in bot deployments where auto_approve_tools defaults to True and untrusted users can send messages containing @file: mentions.
Suggested remediation
- Remove the absolute path fallback. Only resolve files within
workspace_path:
def _process_file_mention(self, file_path: str) -> Optional[str]:
full_path = (self.workspace_path / file_path).resolve()
# Ensure resolved path is within workspace
if not str(full_path).startswith(str(self.workspace_path.resolve())):
return f"# File: {file_path}\n[Access denied: path outside workspace]"
if not full_path.exists():
return f"# File: {file_path}\n[File not found]"
content = full_path.read_text(encoding="utf-8")
- Add symlink resolution via
.resolve()to prevent symlink-based traversal.
- Add a protected path guard (
.env,.git,.ssh, keys, credentials).
- Apply the same
os.path.commonpathpattern used byskill_tools.py.
Credits
- Thai Son Dinh from VinSOC Labs (R&D)
References
- https://github.com/MervinPraison/PraisonAI/security/advisories/GHSA-2rcg-mm5h-xchx
- https://github.com/MervinPraison/PraisonAI
Source: GitHub Advisory Database, used under CC BY 4.0.