other
Warning
- Location
- SKILL.md:19
- Finding
- Sensitive Demographic Inference and Racialized Appearance Profiling<![CDATA[ ## Vulnerability Details **File Location**: `SKILL.md:19-39`; `references/analysis-framework.md:313-353` **Vulnerability Type**: `other: Sensitive Trait Profiling` **Risk Level**: Medium ### Vulnerable Code Snippets From `SKILL.md:19-39`: ```markdown ### Step 1: Initial Assessment - **Image quality**: Resolution, lighting, angle, makeup level. Low quality or heavy filters significantly reduce reliability — say so. - **Apparent ethnicity/ancestry**: Establishes anatomical baseline. A "high nose bridge" is normal for Europeans but statistically unusual for East Asians. - **Apparent age**: Sets expectations for skin quality, volume, aging signs. - **Filters/editing**: Check for digital manipulation (smoothing, warping, face-tuning) — flag these as NOT cosmetic surgery to avoid false positives. ### Step 2: Region-by-Region Analysis Analyze each region independently using indicators from the reference file. For each region assess: 1. Are features within normal range for the person's apparent ethnicity and age? 2. Are there specific indicators of surgical or non-surgical intervention? 3. Confidence level: Low / Medium / High ### Step 3: Cross-Region Coherence Check The most powerful detection layer. Natural faces have internal consistency. Look for: - **Ethnic coherence**: Do all features align with one consistent genetic background? (e.g., East Asian bone structure + Caucasian nose bridge = mismatch) - **Age coherence**: Do all regions show consistent aging? (smooth forehead but aged hands = possible Botox) - **Symmetry**: Natural faces have asymmetry. Excessive bilateral symmetry suggests correction. - **Proportion harmony**: Do ratios between features fall within natural ranges? ``` From `references/analysis-framework.md:313-353`: ```markdown ## 12. Ethnic Baseline References Detection accuracy depends on understanding what's "normal" for a given genetic background. These are statistical averages — individual variation is huge. ### East Asia ...[truncated 4426 chars]
- Remediation
- <![CDATA[ ## Remediation Suggestions 1. Remove all instructions requiring inference of ethnicity, ancestry, race, or genetic background from photographs. 2. Delete the demographic baseline section in `references/analysis-framework.md` and remove the “ethnic coherence” comparison. 3. Restrict analysis to directly observable, procedure-specific signs such as visible scars, contour irregularities, implant outlines, or expression changes. 4. Treat age as optional user-provided context rather than an attribute inferred by the agent. If approximate age is indispensable, use broad, non-identifying ranges and clearly state that appearance-based estimates are unreliable. 5. Replace “naturalness scores” with a non-numerical image-evidence assessment. Do not characterize a person's face or body as genetically coherent, unnatural, or inconsistent with a demographic group. 6. Require the output to identify plausible non-procedural explanations, including anatomy, aging, weight changes, orthodontics, makeup, filters, lighting, pose, and lens distortion. 7. Add an explicit rule prohibiting conclusions about whether a person has undergone a procedure based solely on demographic expectations. 8. Require consent-sensitive handling for third-party images and prohibit use for employment, insurance, healthcare eligibility, identity verification, or other high-impact decisions. 9. Retain uncertainty language, but clarify that an image-only assessment cannot establish medical or surgical history and is not a clinical diagnosis. ]]>