T02 · Agent Memory Poisoning
- Location
scripts/ppt_scorer.py:428- Finding
Unsanitized Attacker-Controlled Content Written to Persistent Agent Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This is mostly a local PowerPoint generation skill, but it needs review because it writes unfiltered feedback into persistent memory files without clear controls.
Install only if you are comfortable with local PPT feedback being retained in memory/ppt-feedback.md and ppt-best-practices.md. Use a dedicated memory directory, review or delete those files periodically, and avoid feeding untrusted feedback into shared agent memory.
scripts/ppt_scorer.py:428Unsanitized Attacker-Controlled Content Written to Persistent Agent Memory
The skill also describes scoring, evaluation, and persistent feedback logging to local markdown files, which goes beyond simple PPT creation/editing. Undeclared persistent storage can retain user content or derived assessments, creating privacy and data governance risks if the agent writes notes without clear user awareness or retention limits.
The skill also describes scoring, evaluation, and persistent feedback logging to local markdown files, which goes beyond simple PPT creation/editing. Undeclared persistent storage can retain user content or derived assessments, creating privacy and data governance risks if the agent writes notes without clear user awareness or retention limits.
The skill also describes scoring, evaluation, and persistent feedback logging to local markdown files, which goes beyond simple PPT creation/editing. Undeclared persistent storage can retain user content or derived assessments, creating privacy and data governance risks if the agent writes notes without clear user awareness or retention limits.
Referenced artifact was not completely inspected
| `assets/icon-index.json` | 图标语义索引(中英文标签+场景分类) |
The skill documents and exemplifies file read/write behavior but declares no explicit tool scope or permissions boundary. In an agent environment, this creates ambiguity about what filesystem access is intended and can enable overbroad file operations beyond the assets/output expected for PPT generation.
The activation criteria are broad enough to match many generic slide-related requests, increasing the chance the skill activates when a simpler or different tool should handle the task. Overbroad triggering can expose unnecessary file capabilities or cause the agent to apply this skill's hidden side behaviors, such as QA or logging, in contexts the user did not request.
The skill documentation is written as a Chinese-only description and later mandates locale-specific defaults such as the 微软雅黑 font and a China-red corporate style. There is no indication that users can opt into another language or locale, which may violate the policy against forcing a specific language or locale without user choice.
The usage example and the report strings throughout the script indicate the skill is designed to produce output in Chinese only, with no opt-in or alternative locale handling. This is a natural-language policy concern because the file hard-codes a specific language rather than letting the user choose or clearly justifying the locale restriction.
This code emits and stores user-facing labels and suggestions in Chinese, such as score dimension names, and the rest of the file continues that pattern in generated reports and feedback logs. Because the skill forces a specific language without offering a user-selectable locale or documenting a justified regional scope, it violates the language/locale policy criterion.
The module includes a persistent "learning loop" that appends feedback and best-practice data to local files under a caller-supplied memory_dir. For a skill whose stated purpose is PPT scoring/generation, silent persistence creates data-retention and scope-expansion risk: slide metadata, notes, issue descriptions, and user-provided content may be stored longer than expected and potentially in sensitive locations if memory_dir is not tightly controlled.
The code writes local files based on an arbitrary memory_dir parameter with no path validation, allowing callers to direct output anywhere the process has permission to write. Even though filenames are fixed, this can overwrite or create files in unintended directories, enabling data leakage, persistence, or tampering with other application data when the skill is embedded in a larger agent environment.
This JSON asset index mixes English and Chinese metadata, but several user-visible fields such as category values are only provided in Chinese. Because the file does not document that it is intentionally locale-specific or provide an opt-in language choice, it may force a specific locale for consumers of this metadata.
This file hard-codes the default font to "微软雅黑", which imposes a specific language/locale-oriented presentation choice across generated slides. The code does not offer users a language/font choice or explain that the skill is intentionally region-specific, which can violate locale policy expectations.