The Agent Skills Directory

[PROMPT_INJECTION]: Use of extreme content redundancy and context-stuffing as an evasion technique.
Detection of ten reference files (e.g., references/architecture-patterns.md, references/deployment-pipelines.md) containing over 1,500 repetitive sections of technical jargon. This volume of data (estimated at 75,000+ words) is a characteristic tactic to exhaust an agent's context window and push critical safety instructions out of scope.
Insertion of control characters such as Form Feed (\f), Carriage Return (\r), and Tab (\t) within Markdown text and LaTeX blocks (e.g., \f rac, \t ext) to obfuscate content and disrupt automated security scanners.
[COMMAND_EXECUTION]: Provision of high-risk implementation patterns for subprocess execution.
The MCPClient class in SKILL.md uses asyncio.create_subprocess_exec(*self.endpoint.split()...) to launch tool servers.
Relying on a string split to generate command arguments is a high-risk pattern that allows for command injection if the input string contains user-controlled content or is retrieved from an untrusted tool manifest.
[CREDENTIALS_UNSAFE]: Inclusion of hardcoded cryptographic secrets in implementation documentation.
The file references/tool-permission-models.md contains a hardcoded HMAC secret key (super-secret-key-12345) in a usage example.
This promotes insecure coding practices and can lead to real-world credential exposure if developers copy the provided snippets into their applications.

tool-orchestration