If you discover a security issue in ctxbudgeter, please do not open a public GitHub issue.

Instead, email karandey3@outlook.com with:

• A description of the issue
• Steps to reproduce
• Affected versions
• Suggested mitigation (if any)

You should receive a response within 5 business days. We aim to release a patch within 30 days for confirmed high-severity issues.

ctxbudgeter is a local-first library. The core never calls an external API. Threats we care about:

• Untrusted YAML pack specs. load_pack() uses yaml.safe_load. Don't load specs from untrusted sources without review.
• Untrusted loaders. Custom loaders run with the same privileges as your process. Treat loader code like any other dependency.
• Sensitivity leakage. Items tagged sensitivity="secret" are flagged in reports and can be enforced via secret_policy. In production, use refuse or redact — never allow. The default is warn, which surfaces the leak but does not prevent it.
• Prompt injection is not addressed by ctxbudgeter. We pack and ship the content you give us; we don't sanitize for injection. Pair this library with a separate input-validation layer for user-supplied content.

• Vulnerabilities in optional dependencies (anthropic, openai, langchain-core, requests, tiktoken, PyYAML). Report those upstream.
• Misconfigured secret_policy: allow in production.
• Loader code you wrote.