Back to skill

Security audit

UE 工程体检

Security checks for vulnerabilities and agentic risk

Overview

This audit tool needs review because baseline output can overwrite project files and its comparisons can miss certain file changes.

Use fresh baseline destinations outside the project, check resolved paths and keep backups: existing files can be overwritten without confirmation. Independently verify important comparisons involving filenames that differ only by case. Enable PIE only deliberately on trusted projects because gameplay can have side effects. Use a trusted local bridge; its implementation was unavailable for review. Scanner coverage gaps were investigated and do not indicate a missing executable payload.

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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (9)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The calls to call_stateful("pie_start", ...) and call_stateful("pie_stop", ...) materially exceed the description's read-only scope. Although explicitly opt-in and designed not to save assets, this check changes editor runtime state and runs project logic. Requiring a local bridge and attempting cleanup do not make it read-only. The supplied code does not directly demonstrate asset writes, automatic fixes, or compilation; the mismatch is the undeclared stateful runtime test.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The baseline-output portion identifies a genuine boundary gap: the CLI validates that report output is outside the project but passes --save-baseline directly to the writer without the equivalent check. A project-local destination can therefore violate the promised no-project-write boundary and potentially overwrite project data; the separate PIE behavior is explicitly disclosed and opt-in.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

Most of the supplied code supports the declared auditing purpose. The exception is call_stateful(), which permits pie_start and pie_stop without passing through the read-only command allowlist. Although this method requires an explicit call and does not itself demonstrate project-file writes or compilation, it provides runtime state-changing behavior beyond the description's read-only inspection scope. The evidence-generation functions appropriately disclose observation limits and are not themselves a mismatch.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, network, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The skill promises read-only inspection but implements stateful Play-In-Editor execution.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

At L079 the check invokes call_stateful("pie_start"), executing the project's gameplay rather than merely inspecting its structure or existing logs; L126 invokes call_stateful("pie_stop"), changing editor state again. Although explicitly opt-in and accurately documented at L001–L004, this capability exceeds the manifest's read-only scope, and not directly saving assets does not make gameplay execution read-only.

Content

No source excerpt is available for this finding.

The CLI exposes a stateful PIE smoke test despite the manifest's read-only inspection scope.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

At L183-L189, --pie explicitly enables a stateful runtime smoke test, and L410 dispatches the PIE check. Opt-in and the promise not to save assets do not make starting gameplay read-only; this goes beyond the manifest's stated inspection-only behavior.

Content

No source excerpt is available for this finding.

The collision-handling comment promises an unknown result, but case-colliding files are silently discarded.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The case-folded map silently drops later files whose names differ only by case, and the diff never records the promised collision uncertainty. On case-sensitive filesystems, changes to a discarded file can evade baseline comparison and produce a falsely clean audit, undermining this tool's verification purpose; Windows-first usage reduces but does not eliminate exposure.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · src/ue_audit/project.py (reported line 159)May include surrounding context.

python
def _run_git(root: Path, args: list[str], timeout: float = 8.0) -> subprocess.CompletedProcess[str] | None:
    try:
        return subprocess.run(
            ["git", "-C", str(root), *args],
            capture_output=True,
            text=True,

The read-only bridge wrapper implements state-changing PIE start and stop commands.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

L61-L65 forwards stateful commands to the editor bridge, with L44 permitting pie_start and pie_stop. These operations change editor runtime state rather than merely inspect project structure, exceeding the manifest's read-only scope even though the method documents them as an explicit opt-in.

Content

No source excerpt is available for this finding.