One key, one point of failure
Most AI agents authenticate with a single API key that grants full access to whatever tools the agent uses. The OWASP Top 10 for LLM Applications flags excessive agency as a critical risk. If that key is leaked, stolen, or extracted, the attacker has exactly the same capabilities as the agent. No secondary check, no approval gate.
The financial industry solved this decades ago with dual authorization for wire transfers. AI agents have not caught up.
The principle
No single entity can authorize an action alone. Multiple parties must independently approve before the action proceeds. For AI agents, the parties might include:
- The agent itself: requests authorization.
- A policy engine: evaluates against defined rules.
- A human operator: reviews high-risk actions.
A 2-of-3 threshold means routine actions proceed automatically (agent + policy engine), while high-risk actions require human approval. This is not about slowing agents down. It ensures the highest-risk actions get scrutiny while routine operations run at machine speed.
EU AI Act Article 14
Article 14 requires that high-risk AI systems allow effective human oversight, including the ability to intervene in or interrupt operation. If your agent has no approval gate, there is no intervention point before a high-risk action is taken. Multi-party approval is the technical mechanism that makes Article 14 compliance practical.
How Asqav implements it
import asqav
asqav.init(api_key="sk_...")
agent = asqav.Agent.create("trading-agent")
# 2 of 3 key shares must approve before this agent signs
config = asqav.create_signing_group(
agent.agent_id,
min_approvals=2,
total_shares=3,
)A signing group splits an agent's signing authority into key shares. You hand the shares to the approving parties, such as a policy engine and a human reviewer, and a gated action only gets signed once the approval threshold is met.
Different groups for different risk levels:
- Low-risk (1-of-1): Agent self-signs. Full audit trail.
- Medium-risk (2-of-3): Agent plus policy engine.
- High-risk (3-of-3): All three must approve.
How this reduces risk
Compromised agent: Attacker has the key but cannot produce the policy engine's or human's signature. Attack is contained.
Data exfiltration: Policy engine flags large data exports. Human reviewer is notified. Exfiltration caught before data leaves.
Unauthorized PII export: Policy engine detects PII fields, requires human approval. Human denies the request.
Every signature is independently recorded. Your audit trail shows who approved, when, and whether anyone rejected. Under DORA and the NIST AI Risk Management Framework, when a regulator asks "who authorized this?", the answer is cryptographic proof, not "the agent did it automatically."
Multi-party approval is an Enterprise feature. Identify your high-risk agent actions, then read the signing groups guide for setup details.