PIIMiddleware Fails To Detect Email, IP, And MAC Addresses Adjacent To Non-ASCII Text Due To Unicode-Aware \B

Python's regex \b is Unicode-aware by default. It matches between a word character and a non-word character, where word characters include all Unicode letters. Patterns anchored with \b therefore fail to start or end when PII is directly adjacent to non-ASCII text. The email, IP, and MAC address detectors all use \b at both ends; the URL detector does not, which is why URL detection works across scripts.

Run the following Python code with langchain 1.3.18:\n```python\nfrom langchain.agents.middleware._redaction import detect_email, detect_ip, detect_mac_address\nprint(detect_email('联系alice@example.com')) # []\nprint(detect_email('Contact alice@example.com')) # ['alice@example.com']\nprint(detect_ip('почта192.168.1.100')) # []\nprint(detect_mac_address('café00:1A:2B:3C:4D:5E')) # []\n```\nThe ASCII-prefixed cases return matches, while non-ASCII-prefixed cases return empty lists.

This change expands detection scope: previously hidden PII next to non-ASCII characters will now be redacted/blocked, which is the intended security improvement. Review any application logic that relies on the old Unicode boundary semantics, especially for mixed-script strings. The credit_card detector still uses \b and remains vulnerable; do not assume full PII coverage after this fix. Rollback: revert the boundary constants to \b in the affected functions. Test across Python versions, as lookbehind assertions with fixed-width alternatives are safe in Python 3.6+. Differential fuzzing reported by the issue author shows no loss of ASCII detection, but run your own regression tests.