Back to skill

Security audit

taichi

Security checks for vulnerabilities and agentic risk

Overview

This is a real multi-agent framework, but it grants unsafe shell execution and performs broad local install/uninstall actions that users should review before installing.

Install only in an isolated development environment with a local, protected Redis instance. Review and patch the subprocess executor before using real data, avoid running the installer with sudo unless you intentionally want host package changes, and do not run uninstall with TAICHI_WORKSPACE pointing outside this framework’s own workspace.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
Findings (4)

T09 · Insecure Skill Coding Practices

Error
Location
taichi-framework/core/skills/skill_executor.py:35
Finding

Untrusted user requests are executed through a shell

Content
View full analysis
dict: return { "nodes": [ { "id": "node1", "description": f"处理: {text}", "depends_on": [], "complexity": 1, } ], "entry_points": ["node1"], "exit_points": ["node1"], } ``` `taichi-framework/core/agents/drafter_agent.py:17-25`: ```python tasks = [] for node in dag.get("nodes", []): tasks.append( { "task_id": node["id"], "skill": "bash_executor", "params": {"command": f"echo 'Executing: {node['description']}'"}, "expected_output": {"stdout": "string"}, "timeout": 30, "retry": 2, } ) ``` `taichi-framework/core/agents/validator_agent.py:14-29`: ```python if message["type"] == "draft.complete": draft = message["payload"] correlation_id = message.get("correlation_id") tasks = draft.get("tasks", []) if not tasks: result = { "status": "rejected", "reasons": ["No tas ...[truncated 3160 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
taichi-framework/core/worker/distributed/distributed_bus.py:25
Finding

Unauthenticated Redis task messages can reach command execution workers

Content
View full analysis
dict: """Execute a single task via SkillExecutor.""" skill_name = task.get("skill", "bash_executor") params = task.get("params", {}) return await self.skill_executor.execute(ski ...[truncated 4516 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
uninstall.sh:34
Finding

Environment-controlled workspace path permits arbitrary recursive deletion

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
install.sh:31
Finding

Root installer uses floating Python dependencies and performs privileged package changes

Content
View full analysis
/dev/null; then echo "[1/4] Redis 未安装,正在安装..." if command -v apt-get &> /dev/null; then sudo apt-get update -qq && sudo apt-get install -y redis-server > /dev/null 2>&1 ``` `install.sh:31-34`: ```bash source "$TAICHI_DIR/venv/bin/activate" pip install --upgrade pip -q pip install -q pyyaml redis click python-dotenv echo " 依赖安装完成" ``` ### Technical Analysis The root-level installer installs Python dependencies by name without version constraints or cryptographic hashes, despite the framework containing a version-pinned `requirements.txt`. The effective dependency set can therefore change between installations based on mutable package-index state. The script also upgrades pip without a version or hash constraint and performs automatic system package-manager operations through `sudo`. No compromised package was identified in the reviewed project, but this installation pattern unnecessarily increases exposure to compromised releases, package-index manipulation, or unsafe index configuration. It also combines application setup with privileged system modification. ### Attack Path 1. A user invokes the documented root installer. 2. The script obtains elevated privileges to update package metadata and install Redis. 3. It upgrades pip from the configured package index without a fixed version or hash. 4. It resolves four Python packages using unconstrained names. 5. If the package source, resolver configuration, or a newly published dependency version is compromised, unreviewed code is installed into the environment and later imported by the framework. ### Impact Assessment A compromised Python dependency can execute code during installation or whenever imported, with the p ...[truncated 252 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
Findings (51)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This finding indicates the skill may stop Redis, delete the Python virtual environment, and remove workspace runtime data despite presenting itself as a workflow framework. In context, that makes the skill significantly more dangerous because it can disrupt services, destroy local state, and sabotage the environment under the guise of normal task orchestration.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This finding indicates the skill may stop Redis, delete the Python virtual environment, and remove workspace runtime data despite presenting itself as a workflow framework. In context, that makes the skill significantly more dangerous because it can disrupt services, destroy local state, and sabotage the environment under the guise of normal task orchestration.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This finding indicates the skill may stop Redis, delete the Python virtual environment, and remove workspace runtime data despite presenting itself as a workflow framework. In context, that makes the skill significantly more dangerous because it can disrupt services, destroy local state, and sabotage the environment under the guise of normal task orchestration.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This finding indicates the skill may stop Redis, delete the Python virtual environment, and remove workspace runtime data despite presenting itself as a workflow framework. In context, that makes the skill significantly more dangerous because it can disrupt services, destroy local state, and sabotage the environment under the guise of normal task orchestration.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This finding indicates the skill may stop Redis, delete the Python virtual environment, and remove workspace runtime data despite presenting itself as a workflow framework. In context, that makes the skill significantly more dangerous because it can disrupt services, destroy local state, and sabotage the environment under the guise of normal task orchestration.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This finding indicates the skill may stop Redis, delete the Python virtual environment, and remove workspace runtime data despite presenting itself as a workflow framework. In context, that makes the skill significantly more dangerous because it can disrupt services, destroy local state, and sabotage the environment under the guise of normal task orchestration.

Content

No source excerpt is available for this finding.

Chaining Abuse

High
Category
Tool Misuse
Confidence
75% confidence
Finding

Tool calls are chained to bypass individual safety checks or escalate capabilities beyond what any single tool call would allow.

Content

Scanner excerpt · install.sh (reported line 16)May include surrounding context.

sh
if ! command -v redis-server &> /dev/null; then
    echo "[1/4] Redis 未安装,正在安装..."
    if command -v apt-get &> /dev/null; then
        sudo apt-get update -qq && sudo apt-get install -y redis-server > /dev/null 2>&1
        echo "  Redis 安装完成"
    else
        echo "  警告: 无法自动安装 Redis,请手动安装 redis-server"

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The manifest exposes a general-purpose subprocess executor that accepts user-controlled command input via command: "{{command}}". Even though allowed_commands is present, the executor remains dangerously overbroad for a Taichi orchestration skill because it enables shell-capable execution primitives unrelated to the stated multi-agent coordination purpose and may allow command abuse or unsafe argument injection depending on how command templating is implemented.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The python_script skill allows execution of an arbitrary script path through python3 {{script_path}}, which creates a direct code-execution primitive. In the context of a multi-agent framework, this is especially dangerous because any agent that can invoke the skill may run unintended local code, bypass intended workflow boundaries, and potentially access sensitive files or perform lateral actions through Python.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The executor builds a shell command by directly substituting untrusted params into a template and then runs it with asyncio.create_subprocess_shell, which invokes a shell interpreter. This enables command injection and arbitrary OS command execution if any parameter is attacker-controlled, and the allowed_commands check only validates the first token after shell parsing, which does not prevent metacharacter chaining or other shell expansion attacks.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

Distributed mode hard-codes use of the bash_executor skill and interpolates user-controlled request text directly into a shell command (echo 'Distributed worker {i}: {request}'). Even though this file is an orchestrator, it is explicitly expanding into shell execution, which creates a command-injection path if quoting is broken and normalizes dangerous execution behavior across workers.

Content

No source excerpt is available for this finding.

Known Vulnerable Dependency: click==8.1.7 — 1 advisory(ies): CVE-2026-7246 (Pallets Click, versions 8.3.2 and below, contain a command injection vulnerabili)

High
Category
Supply Chain
Confidence
96% confidence
Finding

The dependency is pinned to click==8.1.7, and the supplied advisory indicates affected versions up to 8.3.2 include a command injection vulnerability. In a multi-agent framework that may invoke CLI commands or pass agent-produced strings into command handlers, a command injection flaw in a command-processing library can materially increase the risk of arbitrary command execution.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

WORKSPACE_DIR is derived from the TAICHI_WORKSPACE environment variable and is later deleted with 'rm -rf' after only checking that it exists as a directory. This allows an external caller to redirect deletion outside the framework directory, potentially wiping arbitrary user-controlled paths such as a home subdirectory or shared project data.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · uninstall.sh (reported line 28)May include surrounding context.

sh
echo "[2/3] 删除 Python 虚拟环境..."
if [ -d "$TAICHI_DIR/venv" ]; then
    rm -rf "$TAICHI_DIR/venv"
    echo "  venv 已删除"
else
    echo "  venv 不存在,跳过"

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
90% confidence
Finding

The skill instructs the agent to execute shell commands, activate a virtual environment, and run a Python orchestrator, but it does not declare any explicit tool scope such as allowed-tools or permissions. This creates an avoidable trust gap where consumers cannot tell in advance that shell and environment access are required, increasing the risk of unintended command execution or broader capability abuse.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This markdown file contains user-facing natural-language instructions almost entirely in Chinese, and it does not indicate that users may choose another language or that the skill is intentionally limited to a Chinese-speaking context. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
96% confidence
Finding

The script invokes sudo to install redis-server, causing host-level changes with elevated privileges during a routine skill installation. Any privileged operation in an installer materially raises risk because it expands the blast radius of mistakes, dependency compromise, or unexpected script behavior.

Content

Scanner excerpt · install.sh (reported line 16)May include surrounding context.

sh
if ! command -v redis-server &> /dev/null; then
    echo "[1/4] Redis 未安装,正在安装..."
    if command -v apt-get &> /dev/null; then
        sudo apt-get update -qq && sudo apt-get install -y redis-server > /dev/null 2>&1
        echo "  Redis 安装完成"
    else
        echo "  警告: 无法自动安装 Redis,请手动安装 redis-server"

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The installer automatically performs a privileged system package installation using sudo and later starts Redis as a background daemon, which are significant side effects not clearly disclosed or gated by explicit user consent. In a skill installer context, silently changing the host system and launching services increases operational and security risk, especially on shared or production machines.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

This markdown file explicitly states that the skill system supports subprocess invocation, which can execute commands on the host, but it does not include any caution about running external commands or their effect on the local system. The document also later provides an uninstall command, yet nowhere warns users that these actions modify or remove local resources.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The manifest example normalizes execution of arbitrary subprocess commands with templated arguments (python3 /path/to/script.py {{arg1}}) while providing only a minimal command allowlist. In an agent skill framework, this is dangerous because users and downstream authors may copy this pattern and pass untrusted input into shell-invoked subprocesses, leading to command injection or arbitrary code execution depending on implementation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The descriptions Execute a bash command and Run a Python script are extremely broad and provide no policy constraints about when these skills may be invoked. In an agentic system, vague triggers increase the chance that planners or other components will select dangerous execution tools for loosely related tasks, expanding the attack surface and making misuse more likely.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The drafter unconditionally assigns the bash_executor skill to every node, giving all planned tasks shell-execution capability regardless of the node’s actual purpose. Even though the current command only echoes the node description, this design normalizes shell access at the planning stage and can become dangerous if downstream components later trust or expand these drafted commands, especially in a multi-agent framework that routes tasks automatically.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The planner constructs node descriptions with the fixed Chinese prefix "处理:" regardless of the user's language or locale. This forces a specific language in generated content without any visible opt-in, choice, or documented regional constraint.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The agent is positioned as a safety and completeness gate before dispatch, but it approves any non-empty task list without validating task structure, authorization, dangerous actions, or policy compliance. In a multi-agent framework, this creates a fail-open control point where malicious or malformed plans can pass directly to execution, undermining the intended trust boundary.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documented behavior claims safety and completeness validation, but the implementation does not perform those checks, which can mislead integrators into trusting this component as a protective control. That mismatch increases security risk because downstream systems may rely on validator approval as evidence that a plan was screened when it was not.

Content

No source excerpt is available for this finding.