6e62b1ad1a
An agent gets an ML-DSA-65 identity and a policy (spending per time window, allowed tools and recipients, which actions need human approval, who may revoke it). Every action it proposes is judged against the policy, signed by the agent and chained; a verifier that does not trust the agent re-runs the policy over the whole ledger. Approvals and revocations are signed by people with their own ML-DSA-65 keys. The ledger of an agent under a policy is one ledger: a second history is a branch, and two branches are a proof of equivocation anyone can check with the public key alone. The README says what the verifier cannot see: entry times are bounded from below only with a witness (anchors or a start time), and someone who sees one branch cannot know of another. Tests: policy 23/23 with the AIP-23 reference verifier (21 run without it), ledger 51/51, approval and revocation 39/39, command line 23/23; negative control 25/25.
6.6 KiB
6.6 KiB
Aere Quantum
Self-hosted post-quantum infrastructure from Aere Network. Each component is a few files with no dependencies: Node.js 24
and the OpenSSL 3.5 it ships with (node:crypto), nothing from a package registry. The one exception is the notarization
command of the verification layer, which needs ethers.
| component | what it does |
|---|---|
pq-gateway/ |
a TLS 1.3 terminating proxy in front of any HTTP service, with the hybrid key exchange X25519MLKEM768: hybrid-only refuses a classical client at the handshake, hybrid-preferred keeps it working; optional client authentication with ML-DSA certificates |
pq-kms/ |
a transit-style key management service where every key is hybrid: X25519 + ML-KEM-768 for encryption, Ed25519 + ML-DSA-65 for signatures (both halves required); versions, rotation, rewrap, data keys, a chained audit log; the root key from the environment or sealed by an HSM through PKCS#11 |
pq-pki/ |
a private certificate authority for ML-DSA (X.509 v3, RFC 9881): root and issuing CAs, leaf certificates, revocation lists, and a strict chain verifier compared against OpenSSL |
crypto-inventory/ |
a cryptographic inventory of source code (JavaScript/TypeScript, Python, Java, Go, PEM blocks, dependency manifests): every use classified by its exposure to a quantum computer, with a migration target, written as a CycloneDX 1.6 CBOM; nothing from the scanned tree is executed, and its cost stays linear on input built to be slow |
verify-layer/ |
an audit-log sidecar for any deployment: entries are AIP-23 envelopes in a hash chain, the runtime adapter records every running Docker or Kubernetes container without any secret value, and the head of the chain can be notarized on Aere Network for post-quantum finality; it says plainly what that proves (the history before a published head) and what it does not (that the host told the truth) |
proof-kinds/ |
the AIP-23 envelope builder the verification layer uses: fourteen proof kinds, one envelope format, digests instead of raw content |
readiness/ |
the post-quantum readiness scanner of a public hostname: real TLS handshakes (hybrid only, hybrid preferred, TLS 1.2), HSTS, the certificate; no connection to an address not proven public; a rate limit per client that X-Forwarded-For cannot bypass |
control-plane/ |
from findings to a finished migration: a prioritized plan from the inventory and the scanner, its execution through the gateway, KMS and PKI with consent per action and a measured proof afterwards, recipes for the servers the products do not touch and a rescan that judges them, a compliance report against NIST IR 8547, the EU roadmap and optionally CNSA 2.0, and a console that checks it all again |
agents/ |
limits an AI agent cannot break unseen: a post-quantum identity (ML-DSA-65), a policy (spending per time window, allowed tools and recipients, which actions need human approval), a signed ledger of every action judged against the policy, approvals and revocation signed by people, and a verifier that re-runs the policy over the whole ledger without trusting the agent; two branches of one ledger are a proof of equivocation anyone can check |
Each component's README says what it is not and what is not measured. No third party has reviewed any of them.
How each is checked
Every component ships its test suite and a negative control: the control plants a real defect in a copy of the code, one at a time, and requires the named test to fail for the named reason; a planting that cannot be applied, or that breaks the build instead of the test, counts as a failure of the control. Results measured on 2026-09-29 (Node.js 24.14.1, OpenSSL 3.5.5):
| component | tests | negative control |
|---|---|---|
| pq-gateway | 33/33 (node proba-pq-gateway.mjs) |
33/33 (bash proba-pq-gateway-control-negativ.sh) |
| pq-kms | 62/62 (node test/proba.mjs); HSM root on SoftHSM2 + OpenSC 20/20 (test/proba-hsm.mjs, Linux); sealed-file trust rules 7/7 (test/proba-hsm-incredere.mjs) |
16/16 (node test/control-negativ.mjs); sealed-file rules 2/2 in this repository (test/control-negativ-hsm-incredere.mjs) |
| pq-pki | 27/27 (node test/proba.mjs), each verdict compared with OpenSSL 3.5 |
22/22 (node test/control-negativ.mjs) |
| crypto-inventory | 37/37 (node test/proba.mjs); cost on hostile input 8/8 linear (node test/proba-timp.mjs) |
18/18 (node test/control-negativ.mjs); cost 3/3 in this repository (node test/control-negativ-timp.mjs; its fourth case compares with version 0.1.0 from the development history and is skipped here) |
| verify-layer | 34/34 with the AIP-23 reference verifier (AERE_VERIFY_PROOF=<verify-proof.mjs from aere-node> node proba-sidecar.mjs); without it 30 run, 4 are reported as skipped and the exit code is 2 |
6/6 in this repository (node control-negativ-sidecar.mjs; its seventh case compares with the version from the development history and is skipped here) |
| proof-kinds | 24/24 with the same verifier (AERE_VERIFY_PROOF=... node proba-proof-kinds.mjs) |
six negative controls inside the test |
| readiness | 6/6 (node proba-adrese-private.mjs: the private-address rules, and a local listener no scan may touch) |
the rate limit and the queue bound are tested where the service runs, not here (its README says so) |
| control-plane | planner 30/30, command line 9/9, execution 30/30 on real products started locally, remediation 33/33 on real TLS servers, compliance report 27/27 (with the AIP-23 verifier), console 8/8 | remediation 7/7, compliance report 3/3 in this repository |
| agents | policy 23/23 with the AIP-23 verifier (without it 21 run, 2 are reported as skipped and the exit code is 2), ledger 51/51, approval and revocation 39/39, command line 23/23 through files and processes only (on Linux and macOS one more test checks the key file mode; not measured here) | 25/25 (node control-negativ-aprobare.mjs) |
Code comments, most function and variable names (also many exported between the files of a component), test names and control
messages are in Romanian, and so are the two command words of the KMS HSM tool (explained in its README). Error codes, error
messages, the HTTP APIs, the inventory's module interface (scan, buildCbom, renderSummary, ...), the agents' module
interface, command line and data (definePolicy, verifyLedger, approve, ...) and the documentation are in English.
Licence
MIT, see LICENSE. Files: 108 (pq-gateway 6, pq-kms 10, pq-pki 6, crypto-inventory 42, verify-layer 8, proof-kinds 3, control-plane 17, agents 10, readiness 4).