Please report security issues privately via GitHub Security Advisories for this repository (or by contacting the maintainers). Do not open a public issue for exploitable vulnerabilities.

• Domain inputs passed to to_ascii / to_unicode should be treated as untrusted. Callers must not assume the empty-string return from to_ascii is the only error channel; prefer the bool overloads when available.
• Percent-decoding, trimming of spaces/tabs/control characters, and host forbidden-code-point checks are the caller's responsibility (as in the WHATWG URL host parser).
• Prefer the full pipeline (to_ascii / to_unicode) over calling Punycode helpers alone on untrusted data.

• Inputs longer than ada::idna::max_domain_input_bytes (see include/ada/idna/limits.h) are rejected by to_ascii and cause to_unicode(in, out) to return false. This bounds heap growth under adversarial input.

• Large Unicode/IDNA tables are stored DEFLATE-compressed and expanded once on first use into a process-lifetime heap buffer (never freed).
• After inflate, a CRC-32 of the uncompressed payload is checked against the value embedded at build time. A mismatch fails initialization; subsequent API calls fail closed.
• Do not dlclose a shared library that owns this buffer while other code may still call into ada::idna.

• The first call that needs tables pays a one-time inflate (~1 ms typical). Steady-state lookups are O(1) multi-stage table access.