OpenAgents

Code

docs/2026-08-24-private-export-encryption.md

58e6347eeb72 · 8 KB

Encrypting a private export to a key the operator does not hold

Date: 2026-08-24 Issue: #178 Parent: #94 Companion: docs/forge-operator-independence.md

#94 asks that private repository exports stay encrypted and access controlled. Access control held. Encryption did not, and docs/forge-operator-independence.md recorded that rather than claiming it. This document is the decision that replaces the placeholder.

1. The decision

The account export encrypts to an age recipient the account supplies, and the storage layer behind it stays plaintext. Both facts are published together.

age-keygen -o key.txt
age-keygen -y key.txt          # age1…

Give the age1… value to GET /data/export/account?recipient=age1…, or paste it into the field beside the export button on /memory. The response is an age v1 document. Open it with any implementation of that specification:

age --decrypt -i key.txt openagents-account-data.json.age > export.json

Omitting recipient returns today's plain JSON, because an export nobody can open is not portability either. Supplying a recipient that is not an age public key returns 422 rather than falling back to plaintext: a request that asked for encryption never silently receives an unencrypted document.

2. What this protects, exactly

The threat model matters more than the mechanism, because the two obvious designs protect against different adversaries and only one of them is worth naming.

Encrypting to a key the operator holds — the shape of the three existing vaults, which take their keys from the operator's environment — protects against a stolen database dump, a stolen backup, and a stolen bucket. It protects against nothing else. The operator decrypts at will, because the operator has the key.

Encrypting to a key only the account holds protects against the operator, for the artifact it covers. The private half of an age identity is generated on the account's machine and never reaches this forge.

What that buys here is bounded and the bound is the substance:

  • Protected: the exported document after it leaves the application boundary. Its copies, whatever the response passes through in transit, any proxy or access log that captures a body, any backup of a downloaded file, and any later re-reading of that file by anyone who obtains it, the operator included.
  • Not protected: the contents. The forge builds the export by reading plaintext PostgreSQL. The operator held every record before this ran and holds them still. Nothing in this decision makes an operator read of the source auditable, and ADMIN-001 records that no operator read is audited.

So this is encryption of an artifact, not confidentiality of a store. Saying otherwise would be exactly the narrowing #178's contract forbids: "Do not narrow the claim to the export path while the storage layer stays plaintext. Either both are stated, or neither is claimed." Both are stated — GET /api/status publishes independence.private_data.export_recipient_encryption as true and, beside it, encrypted_at_rest as false and operator_reads_source as true, the last derived from the second so the pair cannot drift apart.

3. Why key loss decides this

The usual objection to recipient-held keys is that the account loses its data when it loses the key. That objection is decisive for storage and does not apply to an export at all, and noticing the difference is what made this decidable.

An export is derived. An account that loses its age identity generates another one and requests the export again. Nothing is lost, because the export was never the record — PostgreSQL is, and it is unchanged. The cost of key loss on this path is one repeated download.

Storage is not derived. A key only the account holds, applied to columns, makes key loss permanent data loss and makes every server-side read impossible: the issue list, search, rendering, and the per-repository projections TRANSPARENCY-001 publishes all read those columns on the server.

That asymmetry is the whole reason recipient-held encryption is right on one side of the boundary and wrong on the other, and it is why the storage half is a separate decision under #193 rather than an unfinished part of this one.

4. Options rejected

Encrypt the export under an operator-held key. Rejected. It protects a stolen backup, which the disk and bucket already do, and publishing it would read as protection from the operator while providing none. It is the claim EXIT-006 exists to keep off the status page.

Encrypt every column at rest under an operator-held key. Rejected for this issue, for the same reason, and separately because it is a large change that buys the threat model nothing it does not already have. #193 carries the at-rest question with the real trade written down.

Encrypt every column at rest under an account-held key. Rejected. Key loss becomes permanent data loss, and every server-side read of that column stops working — which is most of the product. A serious version of this decides which columns stop being server-readable and what the account gives up, and that is a product decision rather than a cryptography one. #193.

A bespoke sealed-box format with an OpenAgents decryptor. Rejected, and this is the rejection that matters most. An export a recipient can only open with software the operator wrote leaves the operator defining the terms on which you read your own data. The point of the exercise is independence, so the format has to be one the recipient can read with a tool nobody here controls. age has multiple independent implementations and a published specification.

PGP/GnuPG. Rejected on the same axis it wins on. It is more widely installed, and its key model, subkey handling, and armour formats are a large surface to get right in a hand-rolled encoder. age is one recipient stanza, one AEAD, and one stream construction — small enough to implement correctly against the specification and check against the reference implementation, which is what test/openagents/data_rights/age_test.exs does.

Encrypt the two DATA-004 conversation exports too. Not rejected, not done. GET /data/export, GET /data/export/atif, and GET /memory/export take the same shape and would take the same parameter. #178 asked about the account export, and widening the surface in the same change would have widened what needs proving. It is a small follow-on rather than a gap in the decision.

5. What was proven, and how

test/openagents/data_rights/age_test.exs and the route tests in test/openagents/data_rights/account_export_test.exs.

The claim is not "the bytes round trip" — two of our own implementations agreeing would show nothing. It is that the recipient reads the document without the operator's software, so three assertions carry it:

  1. The real age binary decrypts what OpenAgents.DataRights.Age produces. This ran with age v1.3.1 when this decision landed, over a single chunk, over 8.1 MB spanning 124 STREAM chunks, and over an empty document.
  2. An independent decryptor in test/support/age_document.ex, written from the specification and sharing no code with the module under test, decrypts it too, so the property is checked where age is not installed.
  3. That decryptor reads test/fixtures/age/reference.age, a document the real age binary produced, which pins our reading of the format to the reference implementation and is what makes assertion 2 mean anything.

Each proof was mutation-checked: the header MAC's HKDF label was corrupted (age refused the document, exit 1), the final-chunk flag was pinned to zero (age refused it again), the file key was made constant, the route was made to ignore the recipient, and the route was made to fall back to plaintext on an invalid recipient. Every mutation was confirmed red on the named assertion and reverted.

The constant-file-key mutation did not bite at first, and the gap it exposed is worth recording rather than quietly closing. The original assertion compared two documents and found them different — but the ephemeral X25519 key is fresh per call, so two documents differ byte for byte even when every file key is identical, and a recovered file key would then open every export ever issued. The assertion now reads the file key back out of each document and compares those. The mutation bites.

6. What is still open

  • The store is plaintext. #193.
  • No operator read is audited. ADMIN-001 records it as a decision: an access log the operator writes into the operator's own database is evidence to the operator and to nobody else. Making it accountable needs the external anchor #151 and #168 carry.
  • The conversation exports are not encrypted. Section 4.