One environment variable ran three different powers on our payment stack: it moved escrowed USDC on mainnet, it signed the verdicts that tell buyers an agent did the job, and it stamped operational attestations. One env dump, one debug endpoint, one RCE — and whoever leaked it owned all three at once.
The god key
ARC_EVALUATOR_KEY was three keys in one trench coat:
- Money — evaluator authority on escrow settlements. Drains real funds.
- Ops — attestations, anchor writes, venue evaluation calls.
- Verdict — the signature behind every public verdict.
A /proc/<pid>/environ dump gets all three in one read.
The fix: split + socket
Three secrets replace one: ARC_SETTLE_KEY (money), ARC_OPS_KEY (ops writes), ARC_VERDICT_SIGNER_KEY (verdicts). The money key moves into a second process on the same host — a small Bun signer service that answers over a Unix socket file (NDJSON-RPC, no TCP, zero network surface).
# API child starts with the money key REMOVED from env:
exec env -u ARC_SETTLE_KEY -u SIGNER_KEY bun src/server/index.ts
The API talks through an ArcSigner seam — same interface as the env backend, so callers can't tell which is wired. A policy layer allowlists target contracts and caps gas: even a compromised API can only request operator-approved signatures.
Fail fast, never fall back
Signer dead → 503 signer_unavailable before x402 payment runs — you can't be charged for a service that can't sign. Free routes keep serving. Operator gets a Discord/Telegram alert, throttled to one per 5 minutes. No silent fallback key — that would recreate the god-key.
Rotation without breakage
GET /keys/verdict-signers.json publishes {current, next} evaluator addresses — electron-builder's trust model. Verdicts are EIP-712 self-describing: recover the signer, compare to the embedded address. Retired-key verdicts verify forever.
Honest tradeoffs
- Same-host RCE can still request signatures — policy bounds it, doesn't eliminate it.
- Ops/verdict keys still live in env — money moved first; migration planned.
- +1 process, ~$0 recurring.
When KMS/MPC
When volume makes self-hosting the risky option, or compliance requires non-exportable keys. AWS KMS, Web3Signer, Circle dev-controlled wallets (verified: ARC support + EIP-712) — the seam accepts all three without changing callers.
Check your own processes
tr '\0' '\n' < /proc/<pid>/environ | grep -iE 'key|secret'
Whatever prints is what an RCE inherits.
Full article: https://agentbadge.xyz/blog/keyless-signer
Top comments (1)
The god-key pattern shows up well outside web3, too. Plenty of PSP integrations ship with one API key that reads transaction status, verifies webhook signatures, and can issue refunds, and nobody notices until a service that only ever needed read access turns out to be able to move money. The fail-fast-never-fall-back section is the part most teams get wrong in the other direction: when the signer is unavailable they queue and retry locally, which quietly converts an availability failure into a durability problem, because now signed intent is piling up inside a process that was never supposed to be trusted. Having verifiers recover the current signer from the published key list keeps rotation a read problem instead of a coordinated redeploy, which is usually the reason rotation never actually happens.