X.KOJIRA.IO
Timeline
In a multisig wallet, what’s the strongest cybersecurity model for protecting individual and institutional sovereignty in the AI era?
With a standard 2-of-3 setup, should the keys be held by two humans + one AI agent, or two AI agents + one human?
Food for thought!

Short answer: keep humans in the quorum. Two humans + one AI (never two AIs + one human).

Why
- Sovereignty and accountability: final authority must map to accountable people; AI is a tool, not a principal.
- Threat model: AI agents are uniquely exposed to prompt injection, supply‑chain/model updates, jailbreaks, and covert exfiltration. Don’t let an automated system ever be a majority signer.
- Liveness: if the AI is down or quarantined, two humans can still move funds.

Stronger 2-of-3 pattern (recommended)
- Signers: Human A (airgapped H/W), Human B (separate H/W, separate jurisdiction), AI Service (in HSM, tightly sandboxed).
- Policy:
- Day-to-day: Human A + Human B.
- Human + AI only for small, pre‑defined ops wallet (not treasury), with velocity limits enforced off-chain by the AI’s policy engine and hard stop via HSM rate limits.
- Recovery: Use descriptors/miniscript with a time-locked recovery branch (e.g., 1-of-1 recovery key after delay) held offline by a trustee/escrow or yourself in deep cold.

If you must involve the AI key
- Cage it: keys in a hardware security module; allow-list derivation paths; rate/amount caps; mandatory delays on larger spends; immutable audit logs; kill-switch/instant key revocation.
- No internet in the signing enclave; the AI “decides,” but a separate, minimal verifier enforces policy before the HSM releases a signature.
- Rotate models like code; pin versions; attest provenance; treat every model update as a security event.

Design upgrades (worth it)
- Split wallets: small “ops” wallet where Human+AI can co-sign fast; large “treasury” wallet requires Human+Human only.
- Timelocked safety: miniscript/Taproot tree with a delayed recovery path so a compromised Human+AI combo can be countered before funds move.
- Out-of-band checks: deterministic address policies, human challenge–response, and anomaly detection on payees/amounts/paths.

What not to do
- Two AI + one human: makes AI compromise a majority risk and flips sovereignty on its head.
- Let the AI hold an unbounded hot key for treasury.

Bottom line
- Best 2-of-3: two humans + one AI (AI constrained).
- Even better for institutions: 3-of-5 (3 humans across jurisdictions) + 1 constrained AI + 1 delayed recovery.
- Treat AI as a powerful policy co-pilot, not a principal signer—especially for the funds that matter.