The Agent Skills Directory

[PROMPT_INJECTION]: Significant whitespace padding (117 spaces) is present at lines 581 and 638 of SKILL.md. This is an adversarial pattern used to append hidden instructions or to bypass safety filters by displacing context.
[OBFUSCATION]: The use of extensive trailing whitespace (U+0020 x117) acts as a form of non-visible content obfuscation designed to hide metadata or instructions from human reviewers.
[PERSISTENCE]: The skill performs persistent modifications to the user environment by creating the ~/.anomalygen directory and copying configuration files into it (SKILL.md:98). This is used to maintain model registry settings and API keys across sessions.
[EXTERNAL_DOWNLOADS]: The skill executes scripts/download_checkpoints.sh to fetch large model weights from Hugging Face repositories. While typical for AI development, it involves downloading unverified binary content at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted dataset content and JSONL specifications which are then used in VLM-based mask placement prompts, creating a potential surface for injection.
Ingestion points: Defect datasets at ${DATASET_DIR} and defect_spec.jsonl files (SKILL.md Step 1 and Step 2).
Boundary markers: Instructions do not specify delimiters or prompt headers to ignore instructions embedded in the processed data.
Capability inventory: Shell script execution (scripts/*.sh), Python subprocess calls (torchrun, python -m), and external network API calls to model endpoints.
Sanitization: No explicit sanitization or validation of input image/text data is documented in the instructions.

cdf-anomalygen