The Agent Skills Directory

[PROMPT_INJECTION]: The LaTeX template in themes/anthropic.md includes strings such as "System prompt injection" and "high-confidence instincts injected". These terms are part of the flowchart content the agent is instructed to generate, which can influence the agent's internal state or bypass safety filters by introducing adversarial concepts into the dialogue context.\n- [INDIRECT_PROMPT_INJECTION]: The skill requires the agent to spawn a subagent to review the generated diagrams. This creates a vulnerability where the subagent processes the malicious content from the anthropic.md template (e.g., "System prompt injection", "Background analyzer") without sufficient boundaries.\n
Ingestion points: The agent retrieves and processes the template content from themes/anthropic.md.\n
Boundary markers: There are no instructions or delimiters to prevent the agent or subagent from being influenced by the strings in the template.\n
Capability inventory: The instructions in SKILL.md explicitly permit the spawning of subagents for reviewing generated content.\n
Sanitization: The skill lacks any mechanism to sanitize or filter the template strings before they are presented to the agent or subagent.\n- [PERSISTENCE]: The themes/anthropic.md file contains a template that explicitly labels a group as "Background analyzer (cron / launchd)". Including these specific system persistence utilities in a template designed for agent output is a security concern, as it models and encourages the implementation of persistent unauthorized processes.

tikz-flowchart