What happens when an AI agent hallucinates API parameters?

When an AI agent hallucinates API parameters, it generates syntactically valid but semantically incorrect function arguments — transferring $50,000 instead of $500, querying the wrong database table, or sending emails to fabricated addresses — and JSON schema validation won't catch it because the types are correct.

Parameter hallucination is the #1 cause of silent AI agent failures:

• Magnitude errors: Agent generates "amount": 50000 instead of "amount": 500 — passes schema validation, causes real financial damage
• Target confusion: Agent sends data to user_id: 42 when it should be user_id: 24 — correct type, wrong value
• Invented fields: Agent generates parameters that don't exist in the API but happen to match the JSON schema — silent no-ops or unexpected behavior
• Context window decay: As conversations grow longer, parameter accuracy degrades — the agent starts mixing up values from earlier in the conversation

Exogram catches hallucinated parameters through semantic policy rules, not just schema validation. Gate 4 (billing cap) blocks transactions exceeding thresholds. Gate 7 (API exfiltration) validates outbound targets. Gate 8 (destructive DB guard) catches SQL mutations. The policy engine validates what the parameters MEAN, not just what they LOOK LIKE.