Reiki

Security proxy for agent and LLM traffic: isolates keys, filters what leaves, keeps access reliable

View the Project on GitHub ddproxy/reiki

PHI and NPPI plugin

The phi plugin stops protected health information (PHI) and nonpublic personal information (NPPI) from leaving in an outbound request. It checks each request body for structured identifiers. In block mode a request that holds one is rejected before any key is used. In report mode the request goes through and the finding is logged.

It is a safeguard that catches common identifiers. It does not prove a request is free of protected data. Read What it does not cover before relying on it.

Compiled in by the default plugin-phi feature; inactive until a [plugins.phi] table exists.

Configure

[plugins.phi]
mode = "report"                      # "block" or "report"; required
# kinds = ["ssn", "card", "field"]  # default: every kind below except field, email and phone
# upstreams = ["zen", "openrouter"]  # default: all upstreams
Key Default Meaning
mode required block rejects a request that holds an identifier. report logs it and lets the request through.
kinds all but field, email, phone The kinds to look for.
upstreams all Check only requests to these upstreams. A request to any other upstream is not checked.

Start in report, read the log for false positives, then switch to block. For an application that handles patient records, add field to kinds.

Kinds

Kind Matches
ssn 123-45-6789 with a valid area, group and serial. Nine digits in a row, or spaces instead of dashes, count only after a word such as ssn or social security.
card 13 to 19 digits, grouped by single spaces or dashes, with a plausible first digit and a valid Luhn checksum.
routing Nine digits with a valid ABA checksum after routing or aba.
account Six to 17 digits after account number, account no, account # or acct.
dob A date after dob, date of birth, date_of_birth, birth date or born on.
medical_id A token with at least four digits after mrn, member id, subscriber id, policy number, patient id, medicare id or similar, with spaces or underscores.
npi A 10-digit National Provider Identifier with a valid checksum after npi.
icd10 An ICD-10 diagnosis code such as C50.911 after icd, icd-10 or diagnosis code.
column A cell in a table column whose header names an identifying field, such as first_name, address, zip_code or diagnosis. See Tables.
field A field named like an identifying field, with a value that reads as data: "first_name": "Jane", last_name: Doe, zip_code, phone_number, diagnosis, treatment_plan. Finds names and addresses in JSON, YAML and log lines. Off by default, because source code names these fields too.
email An email address. Off by default, because source code and commit metadata carry many.
phone A US phone number written with separators or parentheses. Off by default.

Tables

A header names a column, so the plugin checks every cell under it as that kind of identifier. A bare nine-digit value under ssn is found even though nothing beside it says so. This covers:

A header is read only on the first line of its table: the first line after a blank line or a line without that delimiter. When a header maps to the column kind, which accepts almost any value, at least half of its non-empty cells must be a known column name. A row later in a table, such as alice,address, is therefore not read as a header, and its cells are not reported as columns.

A table without a recognised header, or an INSERT without a column list, is checked only by the keyword and shape rules above. A cell that holds a type name, null or a placeholder is not reported.

Run reiki scan over sample data to see what this plugin flags before you choose a mode.

Behavior

What it does not cover