The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill describes an architecture where the Accept-Language header is parsed and used to generate instructions that are directly injected into the agent's system prompt.
Ingestion points: User-controlled Accept-Language headers are processed in routes/agents.py via get_language_instruction().
Boundary markers: The code examples show simple string concatenation (lang_suffix = f"\n\n{lang_instruction}") without explicit boundary markers or delimiters that would prevent the injected content from interfering with the rest of the system prompt.
Capability inventory: The skill facilitates the modification of system prompts, which can influence agent behavior, response formatting, and output content.
Sanitization: The provided snippets do not show explicit validation or sanitization of the language strings before they are interpolated into the prompt. If the header parsing logic is overly permissive, it could allow a user to inject unexpected instructions into the agent's context.
[Prompt Manipulation Patterns]: The skill provides tools for developers to dynamically alter system prompts at runtime. While this is a standard internationalization pattern, it creates a mechanism where external data influences the core instructions of the agent. Ensuring that the resulting prompts remain consistent with safety guidelines is a recommended practice.

language-injection