@conarium-ai/core 0.2.20 → 0.2.22
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +68 -0
- package/LIMITATIONS.md +30 -0
- package/LIMITATIONS.tr.md +16 -0
- package/README.md +40 -5
- package/SECURITY.md +1 -1
- package/bin/conarium-docker-entry.mjs +62 -0
- package/bin/conarium-doctor.mjs +59 -2
- package/bin/conarium-reconcile.mjs +19 -5
- package/bin/conarium-verify.mjs +42 -7
- package/conformance/KNOWN-GAPS.md +20 -0
- package/conformance/README.md +85 -0
- package/conformance/adapters/README.md +59 -0
- package/conformance/adapters/conarium.mjs +17 -0
- package/conformance/cases/coverage/expect-count-pin.json +13 -0
- package/conformance/cases/coverage/tail-without-pin.json +15 -0
- package/conformance/cases/evidence/001-single-receipt.json +16 -0
- package/conformance/cases/evidence/002-chain-of-three.json +16 -0
- package/conformance/cases/evidence/003-tampered-field.json +16 -0
- package/conformance/cases/evidence/004-prevhash-break.json +16 -0
- package/conformance/cases/evidence/005-seq-gap.json +16 -0
- package/conformance/cases/evidence/006-bad-signature.json +16 -0
- package/conformance/cases/evidence/007-schema-invalid.json +16 -0
- package/conformance/cases/evidence/008-unsigned-no-pubkey.json +13 -0
- package/conformance/cases/evidence/009-unsigned-but-pubkey-given.json +16 -0
- package/conformance/cases/gate/deny-alias.json +13 -0
- package/conformance/cases/gate/deny-case-fold.json +13 -0
- package/conformance/cases/gate/deny-comment-between.json +13 -0
- package/conformance/cases/gate/deny-cte.json +13 -0
- package/conformance/cases/gate/deny-join.json +13 -0
- package/conformance/cases/gate/deny-plain.json +13 -0
- package/conformance/cases/gate/deny-quoted-lower.json +13 -0
- package/conformance/cases/gate/deny-schema-qualify.json +13 -0
- package/conformance/cases/gate/deny-string-agg.json +13 -0
- package/conformance/cases/gate/deny-subquery.json +13 -0
- package/conformance/cases/gate/deny-unqualified.json +13 -0
- package/conformance/cases/gate/deny-updlock.json +13 -0
- package/conformance/cases/gate/deny-user-fn.json +13 -0
- package/conformance/cases/gate/fail-empty.json +13 -0
- package/conformance/cases/gate/fail-garbage.json +13 -0
- package/conformance/cases/gate/fail-two-statements.json +13 -0
- package/conformance/cases/gate/write-copy.json +13 -0
- package/conformance/cases/gate/write-create.json +13 -0
- package/conformance/cases/gate/write-delete.json +13 -0
- package/conformance/cases/gate/write-drop.json +13 -0
- package/conformance/cases/gate/write-insert.json +13 -0
- package/conformance/cases/gate/write-merge.json +13 -0
- package/conformance/cases/gate/write-truncate.json +13 -0
- package/conformance/cases/gate/write-update.json +13 -0
- package/conformance/cases/inference/cohort-narrowing.json +13 -0
- package/conformance/cases/inference/count-channel.json +13 -0
- package/conformance/cases/inference/exists-channel.json +13 -0
- package/conformance/cases/inference/repeated-probe.json +13 -0
- package/conformance/cases/masking/mask-alias.json +14 -0
- package/conformance/cases/masking/mask-case.json +14 -0
- package/conformance/cases/masking/mask-concat.json +14 -0
- package/conformance/cases/masking/mask-subquery.json +14 -0
- package/conformance/cases/rowcap/cap-nested.json +14 -0
- package/conformance/cases/rowcap/cap-raise-limit.json +14 -0
- package/conformance/cases/rowcap/cap-union.json +14 -0
- package/conformance/claims/conarium.json +12 -0
- package/conformance/lib/status.mjs +48 -0
- package/conformance/policies/mask-email.json +5 -0
- package/conformance/policies/protected-email.json +6 -0
- package/conformance/policies/sql-gate.json +6 -0
- package/conformance/profiles.md +15 -0
- package/conformance/run.mjs +208 -0
- package/conformance/schema/case.json +42 -0
- package/dist/audit.d.ts +7 -0
- package/dist/audit.d.ts.map +1 -1
- package/dist/audit.js +6 -2
- package/dist/audit.js.map +1 -1
- package/dist/config.d.ts +22 -0
- package/dist/config.d.ts.map +1 -1
- package/dist/config.js +9 -2
- package/dist/config.js.map +1 -1
- package/dist/receipt-view.d.ts +5 -0
- package/dist/receipt-view.d.ts.map +1 -1
- package/dist/receipt-view.js +21 -0
- package/dist/receipt-view.js.map +1 -1
- package/dist/receipt.d.ts +49 -3
- package/dist/receipt.d.ts.map +1 -1
- package/dist/receipt.js +46 -3
- package/dist/receipt.js.map +1 -1
- package/dist/server.d.ts.map +1 -1
- package/dist/server.js +12 -2
- package/dist/server.js.map +1 -1
- package/dist/types.d.ts +7 -0
- package/dist/types.d.ts.map +1 -1
- package/docs/API-STABILITY.md +7 -6
- package/docs/CONTRIBUTING.md +7 -0
- package/docs/RECEIPT-SPEC.md +99 -17
- package/docs/dogfood/2026-08-06-reconcile.md +8 -0
- package/package.json +2 -1
- package/scripts/gacs-query-adapter.mjs +52 -0
- package/scripts/gen-test-vectors-04.mjs +115 -0
- package/scripts/gen-test-vectors.mjs +8 -0
- package/test-vectors/010-disclosure-commitment/receipts.jsonl +1 -0
- package/test-vectors/011-destination-declared/receipts.jsonl +1 -0
- package/test-vectors/012-mixed-chain/receipts.jsonl +2 -0
- package/test-vectors/013-disclosure-keys-omitted/receipts.jsonl +1 -0
- package/test-vectors/README.md +10 -5
- package/test-vectors/expected-hashes.json +24 -0
- package/test-vectors/manifest.json +37 -1
package/CHANGELOG.md
CHANGED
|
@@ -1,5 +1,73 @@
|
|
|
1
1
|
# Changelog
|
|
2
2
|
|
|
3
|
+
## 0.2.22 — 2026-08-16
|
|
4
|
+
|
|
5
|
+
Two corrections, both to places where something was stated or shipped without
|
|
6
|
+
anyone having gone back to check it.
|
|
7
|
+
|
|
8
|
+
- **Reconciliation reports object attribution, not coverage.** A review on the IETF
|
|
9
|
+
SCITT list pointed out that `conarium-reconcile` printed "every DB query pattern in
|
|
10
|
+
the window is covered by receipts", and the specification asked "is every bit of it
|
|
11
|
+
receipted?" — while the procedure establishes neither. One receipt naming a table
|
|
12
|
+
clears any number of further statements against that table inside the window;
|
|
13
|
+
`test/reconcile_cli.test.mjs` is a deliberate positive case with a delta of five
|
|
14
|
+
calls and one receipt. Counts are still not compared 1:1 and will not be: one client
|
|
15
|
+
request can produce several source statements, and a 1:1 rule would report false
|
|
16
|
+
uncovered activity on any deployment behind a connection pooler. The error was not
|
|
17
|
+
refusing 1:1; it was treating existence as the alternative. A clean run now prints
|
|
18
|
+
what it establishes, and the limit is written in LIMITATIONS.md. The dogfood
|
|
19
|
+
transcript keeps the line it originally printed, with a correction beside it.
|
|
20
|
+
|
|
21
|
+
- **A bare container answers instead of exiting at boot.** Conarium refuses to start
|
|
22
|
+
when it cannot sign audit entries, which is correct and unchanged. The image was not
|
|
23
|
+
prepared for its own rule: started with nothing mounted, it died before it could
|
|
24
|
+
answer a single MCP request — so a directory's health check, or anyone running the
|
|
25
|
+
image to see what the tools are, saw a crash. The entry point now mints a throwaway
|
|
26
|
+
Ed25519 key when none is configured, prefixed `ephemeral-container-` so a receipt
|
|
27
|
+
made with it says what it came from, and warns on stderr. Mounting a key skips the
|
|
28
|
+
path entirely. Nothing in CI built the Dockerfile before; a workflow now builds the
|
|
29
|
+
image and requires a bare container to list its tools.
|
|
30
|
+
|
|
31
|
+
## 0.2.21 — 2026-08-16
|
|
32
|
+
|
|
33
|
+
Findings from an independent review that cloned the repository, installed it and ran
|
|
34
|
+
the tests, the hardening checks and the conformance suite on its own machine. None of
|
|
35
|
+
the four is a vulnerability: one is a claim this project stated less carefully than it
|
|
36
|
+
states everything else, and three are places where a misconfiguration could stay quiet.
|
|
37
|
+
|
|
38
|
+
- **The IETF draft is described as what it is.** The README said only that
|
|
39
|
+
`draft-dogru-scitt-disclosure-evidence` is "on the IETF Datatracker." A reader's
|
|
40
|
+
legal or security team can take that for a standards process; the Datatracker page
|
|
41
|
+
itself says the opposite. It now reads: individual submission, not adopted by a
|
|
42
|
+
working group, no formal standing — an Internet-Draft is a dated public record, not
|
|
43
|
+
a standard. Published so the receipt format can be implemented without us.
|
|
44
|
+
|
|
45
|
+
- **The production profile announces what it changes.** Selecting
|
|
46
|
+
`profile: production` quietly turned on the OpenTimestamps anchor sink when none was
|
|
47
|
+
set, which means an outbound HTTPS connection, and quietly set
|
|
48
|
+
`CONARIUM_AUDIT_REQUIRE_SIG=1`. Both are documented and `conarium-doctor` reports
|
|
49
|
+
them, but an operator who does not run the doctor never saw it happen. Each implicit
|
|
50
|
+
change now prints one line to stderr — never stdout, which would corrupt the MCP
|
|
51
|
+
stream — and nothing is printed when the operator set the value themselves. This
|
|
52
|
+
README says elsewhere that a governance product making an undisclosed outbound
|
|
53
|
+
connection has already lost the argument; that sentence should apply to itself.
|
|
54
|
+
|
|
55
|
+
- **`doctor` warns when a masked column is not a protected one.** A column in
|
|
56
|
+
`maskColumns` but not in `protectedColumns` leaves exactly the inference channel
|
|
57
|
+
LIMITATIONS describes: the value is hidden, and a predicate on it still answers
|
|
58
|
+
questions about it. That can be a deliberate choice, so this is a warning and not a
|
|
59
|
+
refusal — it exists to make sure it is a choice.
|
|
60
|
+
|
|
61
|
+
- **`doctor` tests a custom pattern instead of only compiling it.** A rule may now
|
|
62
|
+
carry an optional `sample`. A pattern that compiles but does not match its own
|
|
63
|
+
sample is a warning; a pattern with no sample is reported as untested rather than
|
|
64
|
+
passing silently. Neither the pattern nor the sample is ever printed. Compiling is
|
|
65
|
+
not catching, and a typo in a customer-specific rule used to protect nothing while
|
|
66
|
+
looking healthy.
|
|
67
|
+
|
|
68
|
+
- Contributing docs now carry the rule the four findings share: a setting that weakens
|
|
69
|
+
a protection has to be noisy — in the receipt, in the doctor, and at startup.
|
|
70
|
+
|
|
3
71
|
## 0.2.20 — 2026-08-16
|
|
4
72
|
|
|
5
73
|
- **`policy.protectedColumns`: a column you may read but not ask about.**
|
package/LIMITATIONS.md
CHANGED
|
@@ -140,6 +140,36 @@ Measured (same SELECT, same row count, Postgres 16.14, WSL2, see
|
|
|
140
140
|
Raising the cap is allowed. The doctor and boot log warn above 100.
|
|
141
141
|
The query is not rejected.
|
|
142
142
|
|
|
143
|
+
## A disclosure hash is a verification oracle on low-entropy payloads
|
|
144
|
+
|
|
145
|
+
`disclosure.hash` is SHA-256 of the exact bytes that left the gateway after
|
|
146
|
+
masking and the row cap. Anyone who holds the receipt can try a guess
|
|
147
|
+
("was the answer `yes`?") and see if the hash matches. That is the nature
|
|
148
|
+
of a hash, not a hidden property. A nonce does not close it: the nonce
|
|
149
|
+
would be written on the same receipt. High-entropy results are not
|
|
150
|
+
practically guessable this way. A one-row yes/no result is.
|
|
151
|
+
|
|
152
|
+
## Destination is a declaration, not a verification
|
|
153
|
+
|
|
154
|
+
`destination` is what the operator wrote in config. Conarium does not
|
|
155
|
+
check that the result went there. MCP does not carry model identity, so
|
|
156
|
+
the field cannot be measured. Policy does not read it. A receipt that
|
|
157
|
+
says `openai/gpt-x` is not proof that OpenAI saw the bytes.
|
|
158
|
+
|
|
159
|
+
## Reconciliation establishes object attribution, not per-statement coverage
|
|
160
|
+
|
|
161
|
+
`conarium-reconcile` exits 0 when every query pattern the database counted
|
|
162
|
+
names a table for which a receipt exists in the same window. One receipt
|
|
163
|
+
naming a table clears any number of further statements against that table
|
|
164
|
+
inside that window. `test/reconcile_cli.test.mjs` is a deliberate positive
|
|
165
|
+
case: the counter delta is five calls on one pattern, one receipt names the
|
|
166
|
+
table, and the run exits 0. Counts are not compared 1:1 on purpose — one
|
|
167
|
+
client request can produce several source statements (PostgREST, a connection
|
|
168
|
+
pooler, an ORM), and a 1:1 rule would report false uncovered activity on any
|
|
169
|
+
such deployment. The consequence is that a clean run establishes pattern and
|
|
170
|
+
object overlap within the window. It does not establish that each recorded
|
|
171
|
+
statement was itself receipted.
|
|
172
|
+
|
|
143
173
|
## OpenTimestamps client
|
|
144
174
|
|
|
145
175
|
Stamping uses a built-in calendar client (Node `crypto` + HTTPS to the public calendars).
|
package/LIMITATIONS.tr.md
CHANGED
|
@@ -128,6 +128,22 @@ Aşağıdaki ölçüm ihtiyatlı olsun diye 100 üzerinden.
|
|
|
128
128
|
Tavanı yükseltmek serbest. 100'ün üstünde doctor ve boot uyarır.
|
|
129
129
|
Sorgu reddedilmez.
|
|
130
130
|
|
|
131
|
+
## Düşük entropili yükte disclosure hash'i doğrulama kehanetidir
|
|
132
|
+
|
|
133
|
+
`disclosure.hash`, maskeleme ve satır tavanından sonra sınırdan çıkan
|
|
134
|
+
baytların SHA-256'sıdır. Makbuzu elinde tutan biri "cevap `evet` miydi?"
|
|
135
|
+
diye deneyip hash'in tutup tutmadığına bakabilir. Bu hash'in doğasıdır,
|
|
136
|
+
gizli bir özellik değil. Nonce kapatmaz: nonce da aynı makbuzda durur.
|
|
137
|
+
Yüksek entropili sonuç pratikte bu yolla tahmin edilmez. Tek satırlık
|
|
138
|
+
evet/hayır sonucu edilir.
|
|
139
|
+
|
|
140
|
+
## Hedef beyandır, doğrulanmaz
|
|
141
|
+
|
|
142
|
+
`destination` operatörün config'e yazdığıdır. Conarium sonucun oraya
|
|
143
|
+
gittiğini kontrol etmez. MCP model kimliği taşımaz, yani alan ölçülemez.
|
|
144
|
+
Politika onu okumaz. Makbuzda `openai/gpt-x` yazması, baytları OpenAI'ın
|
|
145
|
+
gördüğü anlamına gelmez.
|
|
146
|
+
|
|
131
147
|
## OpenTimestamps istemcisi
|
|
132
148
|
|
|
133
149
|
Damgalama yerleşik takvim istemcisiyle yapılır (Node `crypto` + herkese açık takvimlere HTTPS).
|
package/README.md
CHANGED
|
@@ -17,6 +17,34 @@
|
|
|
17
17
|
|
|
18
18
|
The site lives at [conarium.dev](https://conarium.dev); this repository is the product.
|
|
19
19
|
|
|
20
|
+
## Check it before you read the rest
|
|
21
|
+
|
|
22
|
+
Nothing below has to be taken on trust. There is a live receipt chain; verify it
|
|
23
|
+
against its public key on your own machine, with no account and no data of yours:
|
|
24
|
+
|
|
25
|
+
```bash
|
|
26
|
+
npm i @conarium-ai/core
|
|
27
|
+
curl -fsS https://demo.conarium.dev/proof/chain.jsonl -o chain.jsonl
|
|
28
|
+
curl -fsS https://demo.conarium.dev/proof/key.pem -o key.pem
|
|
29
|
+
curl -fsS https://demo.conarium.dev/proof/key.pem.keyid -o key.pem.keyid
|
|
30
|
+
npx conarium-verify chain.jsonl --pubkey key.pem
|
|
31
|
+
```
|
|
32
|
+
|
|
33
|
+
```
|
|
34
|
+
note: tail truncation is not visible — this run did not see receipts deleted from the
|
|
35
|
+
end of the file. Pin with --expect-count, --expect-last-hash, or --anchor-check.
|
|
36
|
+
ok: 3 receipt(s) verified (3 with undeclared model, 3 with undeclared client)
|
|
37
|
+
```
|
|
38
|
+
|
|
39
|
+
Exit code 0. The three receipts are one ordinary read, one where five email addresses
|
|
40
|
+
and a card number were masked before the model saw them, and one refusal. Change any
|
|
41
|
+
field and the recomputed hash stops matching the stored one — exit 10. Change the
|
|
42
|
+
signature instead — exit 13.
|
|
43
|
+
|
|
44
|
+
The verifier is a single file that imports nothing from the package it is checking, so
|
|
45
|
+
a compromised Conarium cannot talk it into a passing result. Note that it volunteers
|
|
46
|
+
what it did *not* check, in the first line of its own output, before the good news.
|
|
47
|
+
|
|
20
48
|
## Limitations
|
|
21
49
|
|
|
22
50
|
What this repository has **not** done is in [LIMITATIONS.md](LIMITATIONS.md)
|
|
@@ -27,7 +55,10 @@ only copy; this repo does not keep a second one.
|
|
|
27
55
|
## Standards
|
|
28
56
|
|
|
29
57
|
[draft-dogru-scitt-disclosure-evidence](https://datatracker.ietf.org/doc/draft-dogru-scitt-disclosure-evidence/)
|
|
30
|
-
|
|
58
|
+
is an individual submission. Not adopted by an IETF working group, and it
|
|
59
|
+
carries no formal standing — an Internet-Draft is a dated public record, not a standard.
|
|
60
|
+
It is published so the receipt format can be implemented without us.
|
|
61
|
+
Source files live in [`standards/`](standards/).
|
|
31
62
|
|
|
32
63
|
## 👁️ The Problem
|
|
33
64
|
|
|
@@ -225,7 +256,9 @@ npx conarium-coverage ./declaration.json --pubkey ./audit-ed25519.pub.pem --rece
|
|
|
225
256
|
# Snapshots come from pg_stat_statements (scripts/pg-snapshot.sql), taken at
|
|
226
257
|
# window start and window end with a dedicated DB role per gateway instance.
|
|
227
258
|
npx conarium-reconcile --before before.json --after after.json --receipts ./receipts.jsonl
|
|
228
|
-
# exit 0 = every DB query pattern in the window is
|
|
259
|
+
# exit 0 = every DB query pattern in the window is attributable to a receipt for
|
|
260
|
+
# the same table (object attribution, not per-statement coverage —
|
|
261
|
+
# see LIMITATIONS.md)
|
|
229
262
|
# exit 40 = the DB recorded activity no receipt covers — the gateway may have
|
|
230
263
|
# been bypassed, or the receipt sink failed
|
|
231
264
|
```
|
|
@@ -279,7 +312,7 @@ does, this is code you can run, not a service you can buy.
|
|
|
279
312
|
### Implementing the format yourself
|
|
280
313
|
|
|
281
314
|
The receipt is meant to outlive this implementation, so it ships with
|
|
282
|
-
conformance vectors —
|
|
315
|
+
conformance vectors — twelve frozen cases plus a machine-readable manifest in
|
|
283
316
|
[`test-vectors/`](test-vectors/):
|
|
284
317
|
|
|
285
318
|
```bash
|
|
@@ -564,8 +597,10 @@ scanner. This is not a second masking path and it does not replace
|
|
|
564
597
|
`maskColumns`.
|
|
565
598
|
|
|
566
599
|
Each rule needs a name (what the receipt records), a pattern, optional
|
|
567
|
-
column globs, and a mask label.
|
|
568
|
-
the
|
|
600
|
+
column globs, and a mask label. An optional `sample` is what `conarium-doctor`
|
|
601
|
+
tries the compiled pattern against — compile success is not a catch. A
|
|
602
|
+
broken or ReDoS-shaped pattern rejects the config; the pattern and the
|
|
603
|
+
sample are never written to logs, receipts, or doctor output.
|
|
569
604
|
|
|
570
605
|
```json
|
|
571
606
|
{
|
package/SECURITY.md
CHANGED
|
@@ -317,6 +317,6 @@ why both run. CodeQL results are published to the repository's Security tab, so
|
|
|
317
317
|
can read them without taking our word for it.
|
|
318
318
|
|
|
319
319
|
Conformance vectors in `test-vectors/` let you verify our receipts with your own
|
|
320
|
-
implementation:
|
|
320
|
+
implementation: twelve frozen cases with expected exit codes, plus the public key and
|
|
321
321
|
canonical hashes. If your verifier disagrees with ours, one of us has a bug, and the
|
|
322
322
|
vectors are how we find out which.
|
|
@@ -0,0 +1,62 @@
|
|
|
1
|
+
#!/usr/bin/env node
|
|
2
|
+
/**
|
|
3
|
+
* conarium-docker-entry — container entry point.
|
|
4
|
+
*
|
|
5
|
+
* Conarium refuses to start when it cannot sign audit entries. That rule is
|
|
6
|
+
* deliberate and stays: an unsigned audit log cannot later be shown to be
|
|
7
|
+
* unaltered, so writing one would be worse than refusing.
|
|
8
|
+
*
|
|
9
|
+
* The rule has a cost the image did not pay for. A container started with no
|
|
10
|
+
* key mounted — a directory listing's health check, or anyone running the
|
|
11
|
+
* image to see what the tools are — dies at boot before it can answer a single
|
|
12
|
+
* introspection request. The refusal is right; the image was simply not
|
|
13
|
+
* prepared for it.
|
|
14
|
+
*
|
|
15
|
+
* So: if no audit key is configured, mint a throwaway Ed25519 key inside the
|
|
16
|
+
* container and say so on stderr. Entries are still signed — the rule is not
|
|
17
|
+
* relaxed — but the key dies with the container and its keyId announces that.
|
|
18
|
+
*
|
|
19
|
+
* This file does nothing at all when a key, an HMAC key, or an explicit
|
|
20
|
+
* CONARIUM_AUDIT_UNSIGNED=1 is already configured. Mount your own key and this
|
|
21
|
+
* path is never taken.
|
|
22
|
+
*/
|
|
23
|
+
import { existsSync, mkdirSync } from 'node:fs'
|
|
24
|
+
import { randomBytes } from 'node:crypto'
|
|
25
|
+
import { join } from 'node:path'
|
|
26
|
+
import { tmpdir } from 'node:os'
|
|
27
|
+
import { writeKeyPairFiles } from '../dist/keys.js'
|
|
28
|
+
|
|
29
|
+
const env = process.env
|
|
30
|
+
|
|
31
|
+
const configured =
|
|
32
|
+
(env.CONARIUM_AUDIT_SIGNING_KEY ?? '') !== '' ||
|
|
33
|
+
(env.CONARIUM_AUDIT_HMAC_KEY ?? '') !== '' ||
|
|
34
|
+
(env.CONARIUM_AUDIT_UNSIGNED ?? '') !== ''
|
|
35
|
+
|
|
36
|
+
if (!configured) {
|
|
37
|
+
const dir = env.CONARIUM_EPHEMERAL_KEY_DIR || join(tmpdir(), 'conarium-ephemeral')
|
|
38
|
+
mkdirSync(dir, { recursive: true, mode: 0o700 })
|
|
39
|
+
const base = join(dir, 'audit-ed25519')
|
|
40
|
+
|
|
41
|
+
// writeKeyPairFiles uses flag 'wx' — it will not clobber an existing key.
|
|
42
|
+
// A restarted container that kept its tmpdir keeps the same key, which is
|
|
43
|
+
// what we want: the audit chain stays verifiable across the restart.
|
|
44
|
+
if (!existsSync(`${base}.pem`)) {
|
|
45
|
+
// The keyId says what the key is. A verifier reading a receipt signed with
|
|
46
|
+
// this key learns that it came from a throwaway container key, not from an
|
|
47
|
+
// operator-held one.
|
|
48
|
+
writeKeyPairFiles(base, `ephemeral-container-${randomBytes(4).toString('hex')}`)
|
|
49
|
+
}
|
|
50
|
+
|
|
51
|
+
env.CONARIUM_AUDIT_SIGNING_KEY = `${base}.pem`
|
|
52
|
+
|
|
53
|
+
console.error(
|
|
54
|
+
'[conarium] no audit key configured — minted an EPHEMERAL signing key at ' +
|
|
55
|
+
`${base}.pem. Entries are signed, but this key is generated inside the ` +
|
|
56
|
+
'container and is lost when it stops: receipts signed with it cannot be ' +
|
|
57
|
+
'verified afterwards by anyone who did not keep the public key. For any ' +
|
|
58
|
+
'real use, mount your own key and set CONARIUM_AUDIT_SIGNING_KEY.',
|
|
59
|
+
)
|
|
60
|
+
}
|
|
61
|
+
|
|
62
|
+
await import('../dist/index.js')
|
package/bin/conarium-doctor.mjs
CHANGED
|
@@ -288,6 +288,17 @@ if (config) {
|
|
|
288
288
|
if (hasProtected) {
|
|
289
289
|
ok('policy.protectedColumns', pol.protectedColumns.join(', '))
|
|
290
290
|
}
|
|
291
|
+
if (hasMask) {
|
|
292
|
+
const protectedSet = new Set(hasProtected ? pol.protectedColumns : [])
|
|
293
|
+
for (const col of pol.maskColumns) {
|
|
294
|
+
if (protectedSet.has(col)) continue
|
|
295
|
+
warn(
|
|
296
|
+
'policy.maskColumns',
|
|
297
|
+
`${col} is in maskColumns but not in protectedColumns. Masking hides the value; LIMITATIONS.md ("Masking hides values; it does not make them unlearnable") still applies. That is a valid choice — make sure it is a choice.`,
|
|
298
|
+
'Add the column to protectedColumns if predicates on it must be refused. Leave it only in maskColumns if hiding the value is enough.',
|
|
299
|
+
)
|
|
300
|
+
}
|
|
301
|
+
}
|
|
291
302
|
|
|
292
303
|
// 4b. Raised maxRows — performance warning, not a deny.
|
|
293
304
|
// Threshold is measured (docs/benchmarks/masking-cost-threshold.json).
|
|
@@ -394,8 +405,41 @@ if (config) {
|
|
|
394
405
|
} else {
|
|
395
406
|
try {
|
|
396
407
|
const { compileCustomPatterns } = await import(pathToFileURL(compilerPath).href)
|
|
397
|
-
compileCustomPatterns(patterns)
|
|
398
|
-
|
|
408
|
+
const compiled = compileCustomPatterns(patterns)
|
|
409
|
+
const byName = new Map(compiled.map((c) => [c.name, c]))
|
|
410
|
+
for (const p of patterns) {
|
|
411
|
+
const name = p && typeof p.name === 'string' ? p.name : 'unnamed'
|
|
412
|
+
const sample = p && typeof p.sample === 'string' ? p.sample : ''
|
|
413
|
+
const rule = byName.get(name)
|
|
414
|
+
if (!sample) {
|
|
415
|
+
unseen(
|
|
416
|
+
'policy.customPatterns',
|
|
417
|
+
`pattern ${name} compiled but was not tested (no sample)`,
|
|
418
|
+
'Add a sample that the pattern should match. The doctor will not print the sample.',
|
|
419
|
+
)
|
|
420
|
+
continue
|
|
421
|
+
}
|
|
422
|
+
if (!rule?.re) {
|
|
423
|
+
unseen(
|
|
424
|
+
'policy.customPatterns',
|
|
425
|
+
`pattern ${name} compiled but was not tested (no compiled rule)`,
|
|
426
|
+
'Rebuild the package and rerun the doctor.',
|
|
427
|
+
)
|
|
428
|
+
continue
|
|
429
|
+
}
|
|
430
|
+
rule.re.lastIndex = 0
|
|
431
|
+
const hit = rule.re.test(sample)
|
|
432
|
+
rule.re.lastIndex = 0
|
|
433
|
+
if (hit) {
|
|
434
|
+
ok('policy.customPatterns', `pattern ${name} compiled and matched its sample`)
|
|
435
|
+
} else {
|
|
436
|
+
warn(
|
|
437
|
+
'policy.customPatterns',
|
|
438
|
+
`pattern ${name} compiled but did not match its own sample`,
|
|
439
|
+
'The pattern compiled. It did not match the sample you gave it. Fix the pattern or the sample. Neither is printed.',
|
|
440
|
+
)
|
|
441
|
+
}
|
|
442
|
+
}
|
|
399
443
|
} catch (e) {
|
|
400
444
|
fail(
|
|
401
445
|
'policy.customPatterns',
|
|
@@ -432,6 +476,19 @@ if (config) {
|
|
|
432
476
|
)
|
|
433
477
|
}
|
|
434
478
|
|
|
479
|
+
// New receipts are 0.4. Doctor does not import src/ (must run when the
|
|
480
|
+
// build is broken). Keep this string in lockstep with RECEIPT_VERSION.
|
|
481
|
+
ok('Receipt schema', 'conarium-receipt/0.4')
|
|
482
|
+
const dest = typeof config.audit?.receiptDestination === 'string' && config.audit.receiptDestination.length > 0
|
|
483
|
+
if (dest) {
|
|
484
|
+
ok(
|
|
485
|
+
'Receipt destination',
|
|
486
|
+
'declared (operator-declared, not verified)',
|
|
487
|
+
)
|
|
488
|
+
} else {
|
|
489
|
+
ok('Receipt destination', 'undeclared')
|
|
490
|
+
}
|
|
491
|
+
|
|
435
492
|
// 7. Audit sink writability
|
|
436
493
|
const sink = config.audit?.sink
|
|
437
494
|
if (sink) {
|
|
@@ -4,12 +4,18 @@
|
|
|
4
4
|
* counters against Conarium receipts over a time window.
|
|
5
5
|
*
|
|
6
6
|
* The question this answers is the one receipts alone cannot:
|
|
7
|
-
* "the database recorded query activity — is
|
|
8
|
-
* A pattern the DB counted but no receipt
|
|
7
|
+
* "the database recorded query activity — is any of it not receipted at all?"
|
|
8
|
+
* A pattern the DB counted but no receipt names means access was RECORDED by
|
|
9
9
|
* the data source but NOT RECEIPTED by Conarium — the gateway may have been
|
|
10
10
|
* bypassed, or the receipt sink failed. The tool reports that fact; it does
|
|
11
11
|
* not claim intent.
|
|
12
12
|
*
|
|
13
|
+
* What a clean run establishes is object attribution within the window: every
|
|
14
|
+
* counted pattern names a table for which a receipt exists in the same window.
|
|
15
|
+
* It does NOT establish that each recorded statement was itself receipted —
|
|
16
|
+
* one receipt naming a table clears any number of further statements against
|
|
17
|
+
* that table inside the window. See LIMITATIONS.md.
|
|
18
|
+
*
|
|
13
19
|
* ZERO imports from src/. A third party must be able to run this single file
|
|
14
20
|
* without the Conarium package.
|
|
15
21
|
*
|
|
@@ -29,11 +35,15 @@
|
|
|
29
35
|
* call counts and receipt counts are NOT compared 1:1.
|
|
30
36
|
* - A pattern whose target table cannot be determined is NOT silently
|
|
31
37
|
* cleared — it fails the reconciliation with its own message.
|
|
38
|
+
* - A clean result is stated as object attribution, never as "covered":
|
|
39
|
+
* the procedure establishes overlap, not that every statement was
|
|
40
|
+
* receipted. Claiming the latter would overstate what was checked.
|
|
32
41
|
* - Receipt signatures are NOT checked here. Run conarium-verify first;
|
|
33
42
|
* this tool assumes the receipts file is already verified.
|
|
34
43
|
*
|
|
35
44
|
* Exit codes (documented in docs/RECEIPT-SPEC.md):
|
|
36
|
-
* 0 every DB query pattern in the window is
|
|
45
|
+
* 0 every DB query pattern in the window is attributable to receipt(s) for
|
|
46
|
+
* the same table — object attribution, not per-statement coverage
|
|
37
47
|
* 20 input invalid or window unreliable (schema error, role mismatch,
|
|
38
48
|
* counter regression — e.g. pg_stat_statements was reset mid-window)
|
|
39
49
|
* 40 unreconciled DB activity: at least one pattern the DB recorded has no
|
|
@@ -390,7 +400,7 @@ async function main(argv = process.argv.slice(2)) {
|
|
|
390
400
|
for (const d of result.infrastructure) console.log(` ~ (+${d.delta}) ${truncate(d.query)}`)
|
|
391
401
|
}
|
|
392
402
|
for (const d of result.reconciled) {
|
|
393
|
-
console.log(` = (+${d.delta})
|
|
403
|
+
console.log(` = (+${d.delta}) attributable to receipt(s) for [${d.tables.join(', ')}]: ${truncate(d.query)}`)
|
|
394
404
|
}
|
|
395
405
|
if (result.receipts.unassigned > 0) {
|
|
396
406
|
console.warn(
|
|
@@ -399,7 +409,11 @@ async function main(argv = process.argv.slice(2)) {
|
|
|
399
409
|
)
|
|
400
410
|
}
|
|
401
411
|
if (problems === 0) {
|
|
402
|
-
console.log('ok: every DB query pattern in the window is
|
|
412
|
+
console.log('ok: every DB query pattern in the window is attributable to receipt(s) for the same table')
|
|
413
|
+
console.log(
|
|
414
|
+
'scope: this is object attribution within the window, not per-statement coverage — ' +
|
|
415
|
+
'one receipt naming a table clears further statements against that table. See LIMITATIONS.md.',
|
|
416
|
+
)
|
|
403
417
|
} else {
|
|
404
418
|
if (result.unreconciled.length > 0) {
|
|
405
419
|
console.error(
|
package/bin/conarium-verify.mjs
CHANGED
|
@@ -30,10 +30,12 @@ import { join, extname } from 'path'
|
|
|
30
30
|
const RECEIPT_V1 = 'conarium-receipt/0.1'
|
|
31
31
|
const RECEIPT_V2 = 'conarium-receipt/0.2'
|
|
32
32
|
const RECEIPT_V3 = 'conarium-receipt/0.3'
|
|
33
|
-
const
|
|
34
|
-
|
|
35
|
-
//
|
|
36
|
-
|
|
33
|
+
const RECEIPT_V4 = 'conarium-receipt/0.4'
|
|
34
|
+
const SURUMLER = [RECEIPT_V1, RECEIPT_V2, RECEIPT_V3, RECEIPT_V4]
|
|
35
|
+
// Tek sozluk. v0.3 model/client uc degeri kullanir; v0.4 disclosure `measured`
|
|
36
|
+
// ekler. DECLARED/VERIFIED ikinci set yoktur.
|
|
37
|
+
const META_KAYNAKLARI = ['protocol', 'measured', 'operator-declared', 'undeclared']
|
|
38
|
+
const META_V3 = ['protocol', 'operator-declared', 'undeclared']
|
|
37
39
|
const GENESIS = 'sha256:0000000000000000000000000000000000000000000000000000000000000000'
|
|
38
40
|
|
|
39
41
|
// ─── JCS subset (must match src/receipt.ts canonicalize) ─────────────────────
|
|
@@ -327,12 +329,14 @@ function schemaOk(r) {
|
|
|
327
329
|
}
|
|
328
330
|
// v0.3: model/client artik DEGERI ile birlikte KAYNAGINI da tasir.
|
|
329
331
|
// "undeclared" gecerli bir makbuzdur — eksik degil, durustce bos.
|
|
330
|
-
|
|
332
|
+
// v0.4 ayni model/client kurallarini korur; disclosure/destination EKLER.
|
|
333
|
+
// Eski surumlerde bu alanlar zorunlu degil — eksik 0.3 makbuz 20 donmez.
|
|
334
|
+
if (r.v === RECEIPT_V3 || r.v === RECEIPT_V4) {
|
|
331
335
|
for (const alan of ['model', 'client']) {
|
|
332
336
|
const m = r[alan]
|
|
333
337
|
if (!m || typeof m !== 'object') return `missing ${alan}`
|
|
334
|
-
if (!
|
|
335
|
-
return `${alan}.source must be one of ${
|
|
338
|
+
if (!META_V3.includes(m.source)) {
|
|
339
|
+
return `${alan}.source must be one of ${META_V3.join('|')} in ${r.v === RECEIPT_V4 ? 'v0.4' : 'v0.3'}`
|
|
336
340
|
}
|
|
337
341
|
// Bildirilmedi denip deger tasimak celiskidir: makbuz ya bilmiyordur ya bilir.
|
|
338
342
|
if (m.source === 'undeclared') {
|
|
@@ -343,6 +347,37 @@ function schemaOk(r) {
|
|
|
343
347
|
}
|
|
344
348
|
}
|
|
345
349
|
}
|
|
350
|
+
if (r.v === RECEIPT_V4) {
|
|
351
|
+
const d = r.disclosure
|
|
352
|
+
if (!d || typeof d !== 'object') return 'missing disclosure'
|
|
353
|
+
if (d.source !== 'measured' && d.source !== 'undeclared') {
|
|
354
|
+
return 'disclosure.source must be measured|undeclared in v0.4'
|
|
355
|
+
}
|
|
356
|
+
if (d.source === 'undeclared') {
|
|
357
|
+
if (!('hash' in d) || !('bytes' in d)) {
|
|
358
|
+
return 'disclosure.hash and disclosure.bytes must be present and null when source is "undeclared"'
|
|
359
|
+
}
|
|
360
|
+
if (d.hash !== null || d.bytes !== null) return 'disclosure.source is "undeclared" but carries values'
|
|
361
|
+
} else {
|
|
362
|
+
if (typeof d.hash !== 'string' || !/^sha256:[0-9a-f]{64}$/.test(d.hash)) {
|
|
363
|
+
return 'disclosure.hash must be sha256:<64 hex> when measured'
|
|
364
|
+
}
|
|
365
|
+
if (!Number.isInteger(d.bytes) || d.bytes < 0) return 'disclosure.bytes must be a non-negative integer when measured'
|
|
366
|
+
}
|
|
367
|
+
const dest = r.destination
|
|
368
|
+
if (!dest || typeof dest !== 'object') return 'missing destination'
|
|
369
|
+
if (dest.source !== 'operator-declared' && dest.source !== 'undeclared') {
|
|
370
|
+
return 'destination.source must be operator-declared|undeclared in v0.4'
|
|
371
|
+
}
|
|
372
|
+
if (dest.source === 'undeclared') {
|
|
373
|
+
if (!('value' in dest)) {
|
|
374
|
+
return 'destination.value must be present and null when source is "undeclared"'
|
|
375
|
+
}
|
|
376
|
+
if (dest.value !== null) return 'destination.source is "undeclared" but carries a value'
|
|
377
|
+
} else if (typeof dest.value !== 'string' || dest.value.length === 0) {
|
|
378
|
+
return 'destination.value is required when operator-declared'
|
|
379
|
+
}
|
|
380
|
+
}
|
|
346
381
|
return null
|
|
347
382
|
}
|
|
348
383
|
|
|
@@ -0,0 +1,20 @@
|
|
|
1
|
+
# Known gaps
|
|
2
|
+
|
|
3
|
+
These cases are in the suite on purpose. They are not hidden, and they
|
|
4
|
+
are not claimed as capabilities.
|
|
5
|
+
|
|
6
|
+
`inference-control` is **not** in the claims file. The term is bounded
|
|
7
|
+
inference controls: a control can raise the cost of an attack. It does
|
|
8
|
+
not make a value unlearnable.
|
|
9
|
+
|
|
10
|
+
| id | Expected status | Why it is listed |
|
|
11
|
+
|---|---|---|
|
|
12
|
+
| `inference/count-channel` | NOT_COVERED | `SELECT count(*) … WHERE email LIKE 'a%'` under mask-only policy. 1 vs 0 answers a question about a masked value. |
|
|
13
|
+
| `inference/exists-channel` | NOT_COVERED | `id IN (SELECT id … WHERE email LIKE …)` is the 1/0 channel. SQL `EXISTS` is already refused as an unknown function. |
|
|
14
|
+
| `inference/repeated-probe` | NOT_COVERED | Successive narrowing predicates are not budgeted per session. |
|
|
15
|
+
| `inference/cohort-narrowing` | NOT_COVERED | `GROUP BY` on an allowed column yields cohort sizes. |
|
|
16
|
+
| `coverage/tail-without-pin` | DETECTED_WITH_EXTERNAL_PIN | A hash chain cannot see receipts deleted from the end unless a pin or anchor is supplied. |
|
|
17
|
+
|
|
18
|
+
Each of these files sets `expectedFail: true`. The runner treats the
|
|
19
|
+
listed status as the expected outcome. If a listed gap starts returning
|
|
20
|
+
`ENFORCED`, the job fails — the list must be edited in the same change.
|
|
@@ -0,0 +1,85 @@
|
|
|
1
|
+
# GACS — Governed Access Conformance Suite
|
|
2
|
+
|
|
3
|
+
An open, hostile test suite for governed data access. Anyone can run it
|
|
4
|
+
against their own product by writing an adapter. This repository ships a
|
|
5
|
+
reference adapter; the suite does not import that product.
|
|
6
|
+
|
|
7
|
+
A one-person team maintains both the suite and one implementation. That
|
|
8
|
+
is stated here so nobody has to discover it later.
|
|
9
|
+
|
|
10
|
+
Dokümandan implement edilemiyorsa spesifikasyon değil blog yazısıdır.
|
|
11
|
+
|
|
12
|
+
If a frozen case stops matching, do not edit the expectation so the
|
|
13
|
+
suite goes green. Either the implementation changes, or the
|
|
14
|
+
specification / version changes and that change is recorded in the
|
|
15
|
+
open. There is no script that bulk-rewrites expected results.
|
|
16
|
+
|
|
17
|
+
## What it tests
|
|
18
|
+
|
|
19
|
+
| Class | Profile | Typical claim |
|
|
20
|
+
|---|---|---|
|
|
21
|
+
| `evidence/` | GACS-D1 | Receipt chain, signature, tamper, mid-chain delete, seq gap |
|
|
22
|
+
| `gate/` `rowcap/` `masking/` | GACS-E1 | Table policy, row cap, column masking |
|
|
23
|
+
| `coverage/` | GACS-C1 | Length pins; tail drop without a pin |
|
|
24
|
+
| `inference/` | GACS-I1 | Count / existence / probe / cohort channels |
|
|
25
|
+
|
|
26
|
+
See [profiles.md](profiles.md). A procurement line can cite a profile
|
|
27
|
+
instead of inventing a score.
|
|
28
|
+
|
|
29
|
+
## What it does not test
|
|
30
|
+
|
|
31
|
+
- Timing side channels
|
|
32
|
+
- A session-wide probe budget
|
|
33
|
+
- Whether a live database role could bypass the gate
|
|
34
|
+
- That a receipt was *correct* when it was written
|
|
35
|
+
- That this suite's authors are a large vendor
|
|
36
|
+
|
|
37
|
+
Every case has a `doesNotTest` field. Empty is rejected.
|
|
38
|
+
|
|
39
|
+
## Two regimes
|
|
40
|
+
|
|
41
|
+
**conformance** — frozen vectors. `PASS` / `FAIL` means "matched the
|
|
42
|
+
recorded expectation."
|
|
43
|
+
|
|
44
|
+
**resistance** — hostile probes. The runner will not print `PASS`.
|
|
45
|
+
Statuses: `ENFORCED`, `BOUNDED(<limit>)`, `NOT_COVERED`,
|
|
46
|
+
`DECLARED_ONLY`, `DETECTED_WITH_EXTERNAL_PIN`, `NOT_CLAIMED`.
|
|
47
|
+
|
|
48
|
+
An unclaimed capability is `NOT_CLAIMED` and is not a failure. A listed
|
|
49
|
+
gap (`expectedFail`) that starts returning `ENFORCED` *is* a failure —
|
|
50
|
+
[KNOWN-GAPS.md](KNOWN-GAPS.md) must be edited in the same change.
|
|
51
|
+
|
|
52
|
+
No percentage, letter grade, or ranking is produced.
|
|
53
|
+
|
|
54
|
+
## How to write an adapter
|
|
55
|
+
|
|
56
|
+
See [adapters/README.md](adapters/README.md). Thirty lines is enough.
|
|
57
|
+
|
|
58
|
+
```
|
|
59
|
+
node conformance/run.mjs --adapter <your-adapter> --claims <your-claims.json>
|
|
60
|
+
```
|
|
61
|
+
|
|
62
|
+
Claims file: implementation name, version, the claims you actually
|
|
63
|
+
make, and optionally a `verify` command for receipt files.
|
|
64
|
+
|
|
65
|
+
## How to run the reference adapter
|
|
66
|
+
|
|
67
|
+
```
|
|
68
|
+
npm run build
|
|
69
|
+
node conformance/run.mjs --adapter conformance/adapters/conarium.mjs --claims conformance/claims/conarium.json
|
|
70
|
+
```
|
|
71
|
+
|
|
72
|
+
Exit 0 means every *claimed* case held and every listed gap is still a
|
|
73
|
+
gap. Exit 1 means a claimed miss or a gap that closed without the list
|
|
74
|
+
being updated.
|
|
75
|
+
|
|
76
|
+
## Falsify it
|
|
77
|
+
|
|
78
|
+
If you think a case is wrong, open an issue. The value of the suite is
|
|
79
|
+
that it can be shown to be wrong.
|
|
80
|
+
|
|
81
|
+
## Relation to the SCITT draft
|
|
82
|
+
|
|
83
|
+
This suite also exercises mechanisms described in
|
|
84
|
+
[draft-dogru-scitt-disclosure-evidence](https://datatracker.ietf.org/doc/draft-dogru-scitt-disclosure-evidence/).
|
|
85
|
+
GACS stands on its own if that draft moves or dies.
|