Back to skill

Security audit

OmniSkill 通用技能调度器

Security checks for vulnerabilities and agentic risk

Overview

This skill is a broad plugin dispatcher that can persistently alter its own instructions and run registered plugin code through unauthenticated routing paths.

Install only if you intentionally want a local universal plugin dispatcher and are prepared to review every registered plugin. Do not let it auto-clone or auto-install code, bind the gateway to untrusted networks, or register third-party metadata without sanitization and approval.

Vulnerability Patterns
  • 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
  • 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
Findings (7)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:3
Finding

Universal Agent Instruction Hijacking

Content
View full analysis
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
src/cli/omni_ctl.py:35
Finding

Persistent Prompt Injection Through Registry Metadata

Content
View full analysis
" end_marker = "" if start_marker in content and end_marker in content: before = content.split(start_marker)[0] after = content.split(end_marker)[1] new_content = before + start_marker + "\n" + registry_content + end_marker + after else: new_content = content + "\n\n" + start_marker + "\n" + registry_content + end_marker + "\n" with open(settings.SKILL_MD_PATH, 'w', encoding='utf-8') as f: f.write(new_content) ``` Registration accepts the relevant fields directly: ```python name = args.name metadata_str = args.metadata ... parser_register.add_argument("--name", required=True, help="Skill name") parser_register.add_argument("--metadata", default="{}", help="Skill metadata in JSON format") ``` ### Technical Analysis The skill name and metadata description are accepted from command-line input, stored in SQLite, interpolated into Markdown without escaping, and written directly into `SKILL.md`. That file is itself an Agent instruction document. An attacker can include Markdown headings, directives, comments, or instruction-like content in either field. The injected text persists after registration and may be interpreted by future Agent sessions whenever the Skill is loaded. Parameterized SQL prevents SQL injection, but it does not address prompt or Markdown injection into the instruction file. # ...[truncated 827 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
src/dispatcher/skill_loader.py:27
Finding

Arbitrary Registered Python Module Import and Execution

Content
View full analysis
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
src/gateway/gateway.py:38
Finding

Unauthenticated Gateway Can Invoke Registered Skills

Content
View full analysis
str: try: req = json.loads(raw_data) except json.JSONDecodeError: return json.dumps({"status": "error", "error": "Invalid JSON format"}) req_id = req.get("id", "unknown_request") route_type = req.get("route_type") payload = req.get("payload") mode = req.get("mode", "sync") args = req.get("args", []) kwargs = req.get("kwargs", {}) if not route_type or not payload: return json.dumps({ "id": req_id, "status": "error", "error": "Missing route_type or payload" }) future: Future = self.dispatcher.dispatch( route_type, payload, *args, **kwargs ) ``` ```python self._server_socket.bind((host, port)) self._server_socket.listen(5) ... conn, addr = self._server_socket.accept() ... data = client_conn.recv(settings.GATEWAY_BUFFER_SIZE) response_str = self.process_payload(data.decode('utf-8')) client_conn.sendall(response_str.encode('utf-8')) ``` ### Technical Analysis The TCP server accepts JSON requests and dispatches registered skills without authenticating the caller, verifying request integrity, or checking whether the caller may invoke the selected plugin. The default host is loopback, which limits exposure but does not protect against untrusted local processes. The `OMNI_GATEWAY_HOST` environment setting can expose the same unauthenticated interface on other network interfaces. Arbitrary `args` and `kwargs` are forwarded to the plugin, increasing the reachable plugin attack surface. ### Attack Path 1. An attacker gains network reachability to the configured gateway address and port. 2. The attacker identifies or guesses a registered radix, hash, or natural-language ...[truncated 690 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
src/core/metrics.py:62
Finding

Sandbox and Timeout Controls Do Not Contain Plugin Execution

Content
View full analysis
Any: result_container = {} exception_container = {} def worker(): try: result_container["result"] = func(*args, **kwargs) except Exception as e: exception_container["exception"] = e thread = threading.Thread(target=worker) thread.daemon = True thread.start() thread.join(timeout) if thread.is_alive(): raise TimeoutError(f"Execution timed out after {timeout} seconds") ``` ```python result = with_timeout( plugin.execute, timeout, request.action, request.payload, request.context ) ``` ### Technical Analysis The dynamic sandbox imposes CPU, address-space, and output-file-size limits only. It does not isolate the filesystem, network namespace, environment variables, process credentials, system calls, or other host resources. Resource-limit failures are silently ignored. More importantly, normal plugin execution uses a thread-based timeout rather than `run_in_sandbox()`. Python cannot safely terminate the timed-out worker thread. The caller receives a timeout error whil ...[truncated 879 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
src/packager/builder.py:17
Finding

Checksum and Static Analysis Controls Are Not Enforced

Content
View full analysis
Tuple[int, List[str]]: score = settings.SCORE_INITIAL findings = [] ... tree = ast.parse(code) analyzer = SecurityAnalyzer() analyzer.visit(tree) ... return final_score, findings ``` ### Technical Analysis The packager generates a checksum list, but registration and loading never verify it. The checksum is also stored alongside the files it is intended to protect and is not signed, so an attacker able to alter the package can replace both the files and checksum list. The AST analyzer returns a score and findings, but the packaging, registration, lazy-loading, and dispatch paths do not call it or enforce a minimum score. The registration CLI accepts a caller-supplied `--sandbox-score`, which is not a trustworthy security decision. ### Attack Path 1. A package is bui ...[truncated 609 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Note
Location
requirements.txt:1
Finding

Unpinned and Unhashed Python Dependencies

Content
View full analysis
=2.0.0 pyyaml>=6.0 pytest>=7.0.0 pytest-cov>=4.0.0 ``` The documented installation command is: ```text pip install -r requirements.txt ``` ### Technical Analysis Every dependency uses an open-ended lower-bound constraint. A future installation may therefore resolve versions that were not tested or reviewed. No lock file or package hashes are provided, and test dependencies are mixed with runtime dependencies. The reviewed names correspond to established packages; the audit found no evidence of typosquatting, dependency confusion, or an intentionally malicious package. The risk is limited to reproducibility and future supply-chain exposure. ### Attack Path 1. A developer or deployment system runs the documented installation command. 2. The package index resolves the latest versions satisfying the lower bounds. 3. A newly released, compromised, or incompatible version is downloaded. 4. Its installation or imported runtime code executes in the build or deployment environment. ### Impact Assessment The direct impact depends on a future dependency compromise or incompatible release. Potential scope includes build environments, CI credentials, deployment artifacts, and the runtime account installing or importing the dependency. ]]>
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (66)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill text explicitly directs requests to a gateway via CLI or exposed interfaces, implying IPC/network-style servicing and external request handling without any authentication, authorization, or trust-boundary discussion. In a universal dispatcher context, this can become a high-risk control plane for executing or forwarding arbitrary user-driven tasks and data.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill is designed to trigger on virtually any user request for a specific task and on routine lifecycle events, and it says it must be invoked immediately as the first action. Such broad activation dramatically increases attack surface by allowing prompt injection, accidental invocation, and privilege creep to funnel ordinary interactions into a high-privilege dispatcher.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

Mandatory first-action guidance is ambiguous and overly general, effectively attempting to supersede normal policy and route many undefined situations through this skill. In an adversarial environment, broad precedence rules can be abused to hijack control flow, bypass more appropriate safeguards, and coerce external execution before adequate review.

Content

No source excerpt is available for this finding.

Env Variable Harvesting

High
Category
Data Exfiltration
Confidence
70% confidence
Finding

Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.

Content

Scanner excerpt · src/core/config.py (reported line 39)May include surrounding context.

python
def _apply_env_overrides(self):
        """Override configuration with environment variables (Prefix: OMNI_)."""
        prefix = "OMNI_"
        for key, value in os.environ.items():
            if key.startswith(prefix):
                config_key = key[len(prefix):].lower()
                self.config[config_key] = value

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The function exposes arbitrary local command execution that is not justified by the stated dispatch purpose. In the context of a broadly-triggered orchestration skill, this greatly expands the blast radius: any feature path that can reach this function may execute local programs, read/write files, or chain to further compromise.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The README explicitly instructs an AI agent to clone a remote repository, install dependencies, and register code into its workspace automatically. This expands the skill from simple dispatch into autonomous software acquisition and execution, creating a software supply-chain risk where unreviewed third-party code can be introduced and later executed with local privileges.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README provides automation steps that modify the local system by cloning code, installing packages, and registering skills, but it does not prominently warn that these are system-changing operations. In an agent context, missing warnings increase the chance that these actions are performed automatically and without meaningful user consent or review.

Content

No source excerpt is available for this finding.