T09 · Insecure Skill Coding Practices
Warning
- Location
- SKILL.md:50
- Finding
- Persistent Plaintext Storage of Sensitive Third-Party Profiles## Vulnerability Details **File Location**: `SKILL.md:50-67`, `SKILL.md:89-94`, `SKILL.md:145-150`, `SKILL.md:166-170`, `templates/SISTER-PROFILE.md:20-41`, `templates/INTERACTION-LOG.md:3-16` **Vulnerability Type**: Plaintext storage and indefinite retention of sensitive personal information **Risk Level**: Medium ### Vulnerable Code `SKILL.md:50-67`: ```markdown ## Data Storage All data is stored **locally on the user's machine only**. No cloud sync. No external transmission. ``` ~/.sister-skill/ └── sisters/ └── [name]/ ├── PROFILE.md # Structured personality profile └── interaction-log.jsonl # Your observation log ``` - **Storage location**: `~/.sister-skill/sisters/` - **Format**: Markdown profiles + JSONL logs (human-readable plain text) - **Cloud sync**: None. Zero external data transmission. - **Deletion**: Delete any profile by removing its folder - **Portability**: All files are plain text, fully portable ``` `SKILL.md:89-94`: ```markdown ### The Unspoken Layer - What she means when she says "I'm fine" (genuinely fine / not fine / testing if you'll push) - Topics she avoids and why - How she shows love without saying it - What she needs but never asks for ``` `SKILL.md:166-170`: ```markdown 1. **Never overwrite** — always add new memories alongside existing ones 2. **Track confidence** — "described 15 times" vs "mentioned once" 3. **Cross-session persistence** — load existing profiles before responding 4. **Timestamp everything** — every memory is logged with date 5. **Evolution tracking** — people change over time, track the evolution ``` `templates/SISTER-PROFILE.md:20-41`: ```markdown ## Emotional Intelligence - **How she gives advice:** - **How she supports during crisis:** - **How she handles HER crisis:** - **Apology style:** - **What "I'm fine" means when she says it:** ## Humor - **Type:** (self-deprecating / observational / storytelling / chaotic / dry wit) - **What makes her laugh:** - **How sh ...[truncated 3183 chars]
- Remediation
- ## Remediation Suggestions 1. **Obtain informed consent** - Require users to confirm that the profiled person has consented before creating persistent records. - Disable persistent profiling by default for real individuals who have not consented. - Require separate confirmation before generating first-person imitations. 2. **Minimize collected data** - Remove fields that solicit avoided topics, crisis behavior, unspoken needs, trauma, health information, finances, or other sensitive inferences. - Store summarized communication preferences rather than raw memories. - Reject or redact secrets and highly sensitive personal information before persistence. 3. **Protect data at rest** - Encrypt profiles and logs using a platform-backed credential or user-provided key. - Create storage directories and files with owner-only permissions, such as directory mode `0700` and file mode `0600`. - Avoid exposing names directly in predictable directory names; use non-identifying internal identifiers where practical. 4. **Introduce retention controls** - Replace indefinite append-only storage with a configurable retention period. - Provide automatic expiration and secure deletion options. - Allow users to inspect and selectively delete individual observations, not only entire directories. 5. **Make persistence opt-in** - Use session-only processing by default. - Display the exact information and destination file before writing. - Require explicit approval for every persistent update involving sensitive information. 6. **Reduce backup and synchronization exposure** - Warn users that home-directory backup or synchronization software may copy the files despite the Skill having no cloud-sync feature. - Document how to exclude the storage directory from unmanaged backups. - Avoid claiming that no data leaves the device unless backup and host-agent behavior can be controlled. 7. **Add privacy validation** - Before each write, ...[truncated 216 chars]