maskprompt 0.1.1

Sub-millisecond PII redaction for prompts before they reach an LLM.

maskprompt

Sub-millisecond PII redaction for prompts before they reach an LLM. Rust core, Python frontend.

The problem

Compliance wants you to scrub PII before it leaves your VPC and hits an external LLM API. Pure-Python regex passes are not nothing, but they cost hundreds of microseconds per prompt at the p99 and they make the security team nervous because the patterns drift.

maskprompt runs the standard PII detectors (email, phone, credit card with Luhn, SSN, IP addresses, AWS keys, GitHub PATs, JWT) plus your own keyword/phrase lists in a single Aho-Corasick + regex pass, and replaces matches with one of four configurable strategies. The whole thing runs in microseconds at typical prompt sizes.

Install

pip install maskprompt

30-second quickstart

from maskprompt import Masker, BuiltinRule, Strategy

masker = Masker(
    builtins=[BuiltinRule.EMAIL, BuiltinRule.CREDIT_CARD, BuiltinRule.US_SSN],
    custom={"customer": ["Acme Corp"]},  # your own labels
)

text = "Email me at [e-mail] about Acme Corp invoice 4111-1111-1111-1111."
result = masker.mask(text, strategy=Strategy.TAG)

print(result.masked)
# Email me at <EMAIL> about <CUSTOMER> invoice <CREDIT_CARD>.

for m in result.matches:
    print(m.kind, m.start, m.end)

Strategies

Strategy	Replacement	Use it when
Strategy.TAG	<EMAIL>	Default. The LLM sees the type but not the value.
Strategy.HASH	<EMAIL:abc12345>	You need to track "the same redacted value showed up again" without recovering it. blake3 over the original, truncated to 8 hex chars.
Strategy.FIXED	███████	Length-preserving for visual cues.
Strategy.REMOVE	(empty)	When even the type is too much information.

Built-in detectors

Rule	Catches
EMAIL	RFC-5322-ish addresses
US_PHONE	US 10-digit and +1 formats
US_SSN	XXX-XX-XXXX
IPV4	dotted quad
IPV6	:: and full forms
CREDIT_CARD	13–19-digit candidates that pass Luhn
AWS_ACCESS_KEY	AKIA… 20-char keys
GITHUB_TOKEN	ghp_/gho_/ghu_/ghr_/ghs_…
JWT	three base64url segments separated by .

Pick the subset you want; unmentioned detectors are off.

Custom keywords

custom={"label": ["needle1", "needle2"]} labels are case-insensitive and match on word-character boundaries. Pass several labels to tag distinct groups ({"customer": [...], "internal_project": [...]}).

License

Dual-licensed under MIT or Apache-2.0 at your option.