ABS Core Docs

Hallucination Shield

Multi-layer logic analysis to prevent agents from executing actions on false premises.

Hallucination Shield

The Hallucination Shield is a critical governance layer that intercepts tool calls before they are executed, verifying if the agent's logic matches reality.

6-Layer Analysis

  1. Pre-flight Check: Validates if the tool exist and arguments match schema.
  2. Context Integrity: Checks if referenced files or data actually exist in the current session.
  3. Intent Extraction: Uses semantic analysis to determine if the agent knows WHY it is calling the tool.
  4. Policy Matching: Executes local WASM policies against the extracted intent.
  5. Shannon Entropy: Detects obfuscated payloads (jailbreaks/injections).
  6. Logical Consistency: Compares the agent's reasoning in the prompt with the tool arguments.

Example Scenario

Agent: "I will delete user_v1.js because it is empty." Hallucination Shield: Intercepts delete_file. Checks user_v1.js. Finds 400 lines of code. Verdict: DENY - Hallucination detected (Resource is not empty).

On this page