The Agent Skills Directory

[PROMPT_INJECTION]: Large-scale technical bloat and signal-to-noise ratio manipulation. The reference files architecture-patterns.md, code-organization.md, deployment-pipelines.md, error-handling.md, performance-optimization.md, security-best-practices.md, state-management.md, and testing-strategies.md contain 150 repetitive sections each (totalling over 1,200 sections) of technical filler content. This includes irrelevant code snippets in Go, Rust, and TypeScript (e.g., TcpStream handling, Circuit Breaker patterns) that are unrelated to the skill's stated purpose of ML NLP. This volume of data appears intended to saturate the agent's context window and obscure the actual logic.
[PROMPT_INJECTION]: Indirect prompt injection vulnerability surface detected.
Ingestion points: The skill ingests untrusted text data in SKILL.md (via the tokenize_function example) and in references/nlp-model-training.md (via the TextClassificationDataset class).
Boundary markers: The instructions do not define or implement delimiters (e.g., XML tags or triple backticks) or explicit 'ignore embedded instructions' warnings for the input text.
Capability inventory: The skill uses model inference and NLP tokenization libraries. While no direct shell execution was found in the scripts, these capabilities can be manipulated by malicious input data to produce biased or harmful outputs.
Sanitization: No validation, escaping, or filtering of the processed text data was identified.

ml-nlp