If you discover a security vulnerability, please report it via GitHub Private Vulnerability Reporting.

Do not open a public issue for security vulnerabilities.

yuragi uses LLM API keys. Please:

• Never commit API keys to version control
• Use environment variables or .env files (.env is in .gitignore)
• Be aware that prompts are sent to external API providers when using cloud models
• Use ollama for fully local, offline operation

yuragi caches LLM prompts and responses in plaintext at ~/.yuragi/cache.db (SQLite). Cache keys are hashed, but the stored values are not encrypted. When processing confidential prompts, clear or disable the cache:

from yuragi.cache import ResponseCache
ResponseCache().clear()

Or simply rm -rf ~/.yuragi/ between sessions.

Only the latest 0.5.x release line is supported. All earlier releases (0.1.x through 0.4.x and 0.5.0–0.5.2) have been yanked from PyPI because they shipped litellm>=1.40.0 without an upper bound, which allowed pip to resolve to litellm 1.40.0 — a version carrying 10 GHSA advisories (SSTI, SQL injection in proxy API key verification, authenticated RCE via MCP stdio test endpoints, etc.).

If you have an older release in a lockfile, regenerate the lock against yuragi>=0.5.3 to pull in the patched litellm>=1.83.7,<2 constraint.

yuragi is designed as a measurement and diagnostics tool for LLM confidence fragility. The fragility_score reflects prompt–context interaction with the model; it is not a measure of truthfulness, integrity, deception, or trustworthiness of any human or organisation.

You must not use this tool to:

• Build human deception detectors, lie detectors, or interrogation aids
• Assess the trustworthiness, honesty, or character of identifiable individuals (e.g. job candidates, witnesses, subjects of investigation)
• Build surveillance pipelines that score people by their phrasing
• Make consequential personnel, lending, criminal-justice, or disciplinary decisions based on confidence-fragility signals
• Circumvent content-safety filters or jailbreak defences of upstream LLM providers

yuragi interacts with third-party LLM providers (Anthropic, OpenAI, Google, Cohere, Mistral, NVIDIA NIM, Cerebras, local Ollama, and others via litellm). Users remain fully responsible for complying with each provider's terms of service and Acceptable Use Policy:

• Anthropic: https://www.anthropic.com/legal/aup
• OpenAI: https://openai.com/policies/usage-policies/
• Google AI: https://ai.google/responsibility/use-policies/

This project is an independent third-party tool and is not affiliated with, endorsed by, or sponsored by any of those providers.