Security audit
Security checks for vulnerabilities and agentic risk
This skill is a local agent-history logging and analysis helper with no hidden exfiltration or destructive behavior, but its logs may contain sensitive data if users provide it.
Install only if you are comfortable with a local tool storing agent interaction history. Configure log and history paths deliberately, avoid passing secrets, tokens, private file contents, or sensitive command output into recordInteraction, and clear or protect the generated log/history files when they are no longer needed.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
},
{
tool: 'read',
path: '/etc/passwd',
success: true,
duration: 0.05
}
The skill explicitly promotes recording tool calls, responses, errors, and audit history but does not warn that these records may contain secrets, personal data, file contents, or command outputs. In an agent context, runtime history often includes highly sensitive material, so normalizing broad logging without minimization or redaction creates a real privacy and data-exposure risk.
The example shows placing an API key directly in configuration without any warning about secret handling, storage, or exposure. Users may copy this pattern into code, commit credentials to source control, or leak them through logs and reports generated by the same skill.
The audit example uses includeAllData: true and elsewhere encourages storing full interaction details, including user content, tool calls, responses, and errors. In an agent skill, this can capture prompts, credentials, proprietary data, filesystem paths, command output, and other sensitive material, creating a significant leakage risk through logs, reports, or exports.
The documentation says reports can be stored in memory, written to log files, and exported, but does not warn about persistence, access control, or privacy implications. Because the reports aggregate interaction history and possibly tool output, exporting them can materially increase the blast radius of any sensitive data captured during operation.
logEvolution() appends full improvement objects to a log file and also adds them to accumulated history, which is later serialized to disk. Even if the logged object is not directly user input, it is derived from runtime behavior and can still encode sensitive operational details or references to prior interactions, creating a durable disclosure path.
The skill persists runtime interaction data to disk automatically in recordInteraction() without any consent, notice, minimization, or retention control beyond a global history cap. Because data is stored verbatim, sensitive prompts, tool outputs, tokens, or user content can be retained locally and exposed through filesystem access, backups, or later inspection.
recordInteraction() stores arbitrary user/agent data into this.history and writes it as plain JSON to disk, creating a straightforward sensitive-data retention channel. In an agent-evolution skill, runtime history is especially likely to contain prompts, responses, error traces, and tool inputs/outputs, so the context makes plaintext logging more dangerous rather than less.
Most of the user-facing documentation, feature descriptions, examples, and changelog notes are written in Chinese, with no indication that other languages are supported or that the language is intentionally region-specific. This can violate a language-choice policy when the skill forces a specific locale without offering user opt-in.
This JavaScript file contains natural-language comments and console output entirely in Chinese, including the visible runtime messages shown to users during execution. Under the policy, forcing a specific language without user opt-in can be a natural-language policy violation when no alternative locale choice or justification is provided.
No suspicious patterns detected.