Skip to main content

​
Policy Engine

The ABS Policy Engine is the core decision-making component of the ABS Kernel. It evaluates every agent action against a set of active policies to determine whether to ALLOW, DENY, or ESCALATE the operation.

​
Architecture

The engine operates on a Fail-Close principle: if a policy cannot be evaluated or an error occurs, the default action is to block.

​
Decision Flow

  1. Event Ingestion: The engine receives an EventEnvelope containing the agent’s intent, payload, and metadata.
  2. Sanitization: Input safeguards remove PII and check for prompt injection.
  3. Policy Evaluation: The active policy for the event type is executed.
  4. Risk Scoring: A risk score (0-100) is calculated based on:
    • Static Rules: Regex matches, keyword blocks.
    • Contextual Analysis: Velocity checks, anomaly detection.
    • Sequence Analysis: Recognition of dangerous workflows.
    • Trust Score: Historical reputation of the agent.
  5. Verdict:
    • ALLOW: Risk score < 30.
    • ESCALATE: Risk score 30-79 (requires human approval).
    • DENY: Risk score >= 80 or explicit block.
  6. Immutable Log: The decision is signed and recorded in the Write-Ahead Log (WAL).

​
Policy Types

​
Static Policies (Scanner Mode)

Designed for high-performance, local execution.
  • Regex: Detect PII (CPFs, Credit Cards, API Keys).
  • Keyword: Block “password”, “secret”, “groselha”.
  • Structure: Validate JSON schemas and file types.

​
Dynamic Policies (Runtime Mode)

Leverage state and LLM intelligence.
  • Velocity: “User X sent > 5 messages in 1 second.”
  • Anomaly: “This code commit style differs from User Y’s profile.”
  • Semantic: “The user is trying to bypass authentication” (Intent detection).

​
Example Policy

export const FINANCIAL_POLICY = {
  name: "financial-safe-transfer",
  description: "Enforces limits on financial transactions",
  evaluate: (proposal, event) => {
    const amount = event.payload.amount;

    // Critical Risk
    if (amount > 50000) {
      return { decision: "DENY", reason: "Amount exceeds hard limit" };
    }

    // High Risk - Human Review
    if (amount > 10000) {
      return { decision: "ESCALATE", reason: "High value transfer" };
    }

    return "ALLOW";
  },
};