The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions are highly prescriptive regarding academic methodology, APA citations, and narrative formatting. No attempts to bypass safety filters, override system instructions, or extract system prompts were found.
[DATA_EXFILTRATION]: The skill uses web_search and web_fetch to gather information from the public internet. It does not attempt to access sensitive local files (e.g., SSH keys, AWS credentials) or exfiltrate private user data to unauthorized domains.
[REMOTE_CODE_EXECUTION]: The skill relies on native platform tools (web_search, web_fetch, sessions_spawn) for its operations. There are no instances of downloading or executing untrusted scripts, shell commands, or external binaries.
[INDIRECT_PROMPT_INJECTION]: The skill has an inherent attack surface for indirect prompt injection because it ingests untrusted data from the web via web_fetch (Phase 3).
Ingestion points: Web content retrieved via web_fetch as defined in SKILL.md and reference/protocol.md.
Boundary markers: Not explicitly defined for the fetched content in the instructions.
Capability inventory: web_search, web_fetch, and sessions_spawn for parallelizing research tasks.
Sanitization: The skill emphasizes academic synthesis and alignment with existing evidence, which serves as a natural reasoning check, but does not specify technical sanitization. Given the research-focused nature and the requirement for user checkpoints in Phase 1, 2, and 4, the risk is managed.
[COMMAND_EXECUTION]: No dangerous shell command execution patterns or privilege escalation attempts were identified. Usage of sub-sessions is restricted to the research methodology defined in the skill documentation.

academic-deep-research