T09 · Insecure Skill Coding Practices
- Location
domains/troubleshoot.json:56- Finding
Unsafe Recommendation to Make a Private Training Dataset Public
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a mostly transparent Solo Hub guide, but it includes unsafe advice that could lead a user to make a private training dataset public without enough warning.
Review private datasets carefully before following troubleshooting advice. Do not make any dataset public unless you have confirmed it contains no secrets, personal data, proprietary content, regulated data, or licensing restrictions, and understand public copies may persist. Enter HF and W&B tokens only in the authenticated Solo Hub UI using the minimum scopes needed.
domains/troubleshoot.json:56Unsafe Recommendation to Make a Private Training Dataset Public
The tutorial explicitly instructs users to paste HuggingFace and Weights & Biases API tokens into the UI but provides no warning about treating those values as secrets, using least-privilege scopes, or avoiding exposure in shared/org contexts. In an interactive agent setting, users may paste credentials into logs, screenshots, recordings, or the wrong interface, increasing the risk of token leakage and unauthorized access to external accounts.
Across the manifest, the file defines a multi-step procedure but never states when the overall skill should or should not be invoked. Under SQP-1 for manifest files, missing specificity on trigger scope is a valid concern because ordinary requests about models, GPUs, datasets, or costs could overlap with this workflow without a narrow activation boundary.
This JSON manifest-like file describes steps such as 'Select LLM as the job type in the fine-tuning wizard' but does not define explicit trigger phrases, scope limits, or exclusion conditions for when this skill should activate. Because SQP-1 applies to manifest files, the lack of specific invocation boundaries is a quality issue that may cause the skill to match broadly to generic fine-tuning requests.
This JSON file defines multiple skill steps and descriptions, but it does not state how or when the skill is invoked, nor does it provide explicit trigger phrases or exclusion conditions. For manifest files, missing trigger specificity can cause unintended activation because the operational scope is left implicit.
The tutorial requires fixed English-centric formats and labels such as owner/dataset-name, username/model-name, and specific token naming conventions in its natural-language instructions. While this may be operationally common, the file does not offer any language or locale choice or note that these identifiers are format constraints rather than user-language requirements.