aissert is an evaluation harness for Claude Code skills. Its main security concerns are data leakage, prompt-injection resistance, and safe plugin packaging.

Security fixes target the latest released 0.x version.

Please do not open a public issue for a vulnerability or data exposure report. Use GitHub private vulnerability reporting if it is enabled for the repository, or contact the maintainer through the GitHub profile linked from the project.

Include:

• affected version or commit SHA;
• a minimal reproduction;
• whether sensitive data may have been exposed;
• logs or artifacts with secrets and proprietary data removed.

Do not put real golden sets, Jira snapshots, Confluence exports, customer data, or proprietary work-system data inside this repository directory.

This includes ignored folders. Directory-source plugin installs copy the whole working tree into the Claude plugin cache, including ignored files. Keep real golden sets outside the repo, for example:

~/golden-sets/<skill>/

The public repository should contain only synthetic fixtures such as golden/example/ and the synthetic canary set.

Evaluated skill outputs and golden snapshots are untrusted input.

aissert mitigates this by:

• keeping judge agents on tools: [];
• requiring the orchestrator to pass content into prompts instead of allowing agents to read files;
• treating malformed agent JSON as a pipeline error;
• computing all metrics in Python, not in an LLM;
• running canary checks to detect judge behavior drift.

Claude release zips are built from an allowlist in scripts/build_claude_plugin_zip.py. Development directories such as tests/, knowledge/, .venv/, .idea/, and local golden-set folders are excluded by construction.

If a new runtime path is needed, add it to the allowlist deliberately and verify the archive contents before release.