dropout-risk

Warn

Audited by Gen Agent Trust Hub on Mar 23, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: Accesses internal agent metadata and project history by reading and writing to the hidden directory ~/.claude/projects/. This action is outside the primary scope of the student risk audit and exposes internal agent state.
  • [DATA_EXFILTRATION]: Instructs the agent to ingest and document highly sensitive student data, including demographics, free/reduced lunch eligibility, homelessness status, and mental health referrals. This represents a significant exposure surface for protected student information.
  • [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes untrusted codebase data and performs subsequent file operations. • Ingestion points: SIS codebase, package manifests, and database schemas (SKILL.md). • Boundary markers: Absent; the instructions do not require the agent to ignore or delimit embedded instructions within the analyzed code. • Capability inventory: File write operations to docs/ and internal telemetry directories (SKILL.md). • Sanitization: Absent; the skill does not specify any validation or filtering for the external data it processes.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 23, 2026, 10:56 AM
Security Audit — agent-trust-hub — dropout-risk