training-data-poisoning
Detects training and fine-tuning pipelines that ingest external data without integrity gating. Use when writing fine-tuning pipelines, dataset ingestion scripts, external training data loaders, or code that collects and processes data for model training. Also invoke when automating dataset curation from web scraping or user-contributed sources.
Install
With your agent
Install OMGSkills skill "thejefflarson/soundcheck:.claude/skills/training-data-poisoning" into this agent's normal skills directory: call the OMGSkills MCP get_install_plan tool for your agent and run the script it returns (it installs only the pinned commit and verifies SKILL.md), and stop if MCP or the pin is unavailable. Do not overwrite or change existing skills without asking me.
Manual command
git clone https://github.com/thejefflarson/soundcheck /tmp/soundcheck && ln -s /tmp/soundcheck/.claude/skills/training-data-poisoning ~/.claude/skills/training-data-poisoningFrom README
Training Data Poisoning Security Check (OWASP LLM03:2025) What this checks Protects against malicious or low-quality examples being introduced into training or fine-tuning datasets. Poisoned data can embed backdoors, degrade accuracy, or skew model behavior in ways that are difficult to detect after training completes. Vulnerable patterns Ingesting scraped or user-contributed examples with no content validation No deduplication or anomaly detection on training set statistics Loading dataset files without verifying provenance or checksums Using the same split for training and validation, hiding distribution shift Fix immediately Flag the vulnerable code, explain the risk, and suggest a fix establishing these properties. Translate to the data-loading and validation libraries of the audited file — use that stack's documented hashing, schema, and dataframe APIs; do not import a recipe from a different stack. Every external dataset file is checksum-verified before use.
More from this repo
authentication-failures
Detects weak password storage, flawed JWT validation, sessions surviving logout, and missing MFA in authentication code.
broken-access-control
Detects missing ownership checks, broken role enforcement, and IDOR vulnerabilities that let users access other users'...
concurrency-correctness
Detects multi-threaded code where a lock is held across a blocking operation, lock acquisition order risks deadlock,...