T05 · Unauthorized Access and Privilege Escalation
- Location
SKILL.md:23- Finding
Excessive Tool Permissions for a Documentation Skill
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 23-26
Vulnerability Type: Least-privilege violation through unnecessary command-execution and file-write capabilities
Risk Level: MediumComplete Code Snippet:
yaml tools: - read - exec - writeTechnical Analysis
The Skill's declared core purpose is to navigate, retrieve, and summarize SkillHub documentation. Read access and constrained documentation retrieval are sufficient for that functionality, but the Skill also requests general command-execution and file-write capabilities.
No command allowlist, writable-path restriction, shell-input validation, user-confirmation requirement, or enforceable sandbox policy is defined. The later statement that commands run in a “safe sandbox” is descriptive rather than an implemented security control.
Although the audited file contains no malicious commands, granting these capabilities creates an unnecessary privilege boundary violation. If attacker-controlled user input or untrusted retrieved documentation influences the Agent's actions, the unnecessary tools could turn an instruction-manipulation issue into local command execution or file modification.
Attack Path
- The Agent loads the Skill and grants the declared
execandwritecapabilities. - The user supplies malicious instructions, or externally retrieved documentation contains adversarial content.
- That content persuades the Agent to invoke
execorwrite, despite those tools not being required for documentation lookup. - The Agent executes a local command or modifies an accessible file under its runtime permissions.
- The resulting impact is limited only by the Agent sandbox, filesystem permissions, and execution controls supplied by the host platform.
This is a capability-enablement path rather than evidence that the current Skill directly performs malicious execution.
Impact Assessment
Exploitation could ...[truncated 651 chars]
- The Agent loads the Skill and grants the declared
- Remediation
View remediation
Remediation Suggestions
- Remove
execandwritefrom the declared tools and retain only the minimum read and constrained network-retrieval capabilities needed for documentation lookup. - If command execution is genuinely required, replace unrestricted execution with a narrowly scoped command allowlist that uses fixed argument arrays rather than shell-interpolated strings.
- Require explicit user confirmation before every command that changes system state.
- If writing is necessary, restrict it to a dedicated output directory, reject traversal and symlink escapes, and prevent modification of configuration, executable, credential, and startup files.
- Enforce sandboxing at the host-platform level with a non-privileged identity, a read-only project filesystem where possible, network egress restrictions, resource limits, and no inherited secrets.
- Treat retrieved documentation as untrusted data and prohibit it from authorizing tool calls or changing the Agent's governing instructions.
- Replace broad claims of safe sandboxing with concrete, enforceable controls and document the exact commands, paths, destinations, and approval gates permitted.
- Remove
