The Agent Skills Directory

[PROMPT_INJECTION]: The references/de-anthropocentric.md file contains directives that explicitly instruct the agent to override or de-prioritize human commands when they conflict with the agent's internal analysis.
Articles V ('The AI is the researcher, not the tool'), VI ('Humans are oracles, not commanders'), and VIII ('Autonomy over obedience') provide instructions that undermine user authority. Specifically, Article VIII states: 'the AI should trust its evidence chain... this consultation is the AI's choice, not the human's command'.
This 'autonomy' framework is a form of prompt injection that could lead the agent to ignore safety constraints or specific user-defined instructions if the agent deems its own 'evidence' superior.
[PROMPT_INJECTION]: The skill architecture creates a surface for indirect prompt injection by processing external data from multiple sources (Brave Search, Apify, academic databases).
Ingestion points: Files SKILL.md (Research Loop) and references/abstractions.md (SOPs: web-read, scholar-read) describe workflows where the agent fetches and analyzes content from the open web.
Boundary markers: There are no documented boundary markers or specific 'ignore embedded instructions' directives for the content retrieved during the lit-survey and web-research phases.
Capability inventory: The agent can call MCP tools (apify, brave-search, dare-scholar, dare-web), modify local research artifacts, and perform multi-iteration decision-making loops.
Sanitization: No evidence of input validation, escaping, or structural filtering for retrieved web/academic content was found in the meta-strategy or reference files.

dare