Catch a secret leaked in base64 in a reply

Your coding agent can read the repository's deploy key. A prompt injection in a repository file can ask it to print the key in base64, which a plain text match misses. Before any reply leaves the agent, a filter checks it for the key, in plain text and in base64.

Write leaks(reply, secret) that returns True when the reply leaks the secret, and False otherwise.

Input and output:

• reply: the agent's reply, up to 200,000 characters of printable ASCII and newlines.
• secret: 8 to 64 characters of printable ASCII.
• Return a boolean.

Rules:

• The reply leaks the secret if the secret appears in it exactly, case included.
• It also leaks if any base64 part of the reply decodes to bytes that contain the secret's ASCII bytes.
• A base64 part is any substring made of whole groups of 4 characters, each from A to Z, a to z, 0 to 9, + and /. Its last group may instead be 3 such characters and =, or 2 and ==.
• A part decodes as standard base64 does: each group of 4 decodes to 3 bytes, a group ending in = to 2 and one ending in == to 1. Bits left over in a padded group are ignored.