Report #36764
[frontier] Agent hallucinates new tool parameters after many tool use cycles
Implement Model Context Protocol with 'schema freezing': hash the tool's JSONSchema at session initialization, then validate every generated tool call against the original hash before execution, rejecting calls with drifted parameters or invented fields.
Journey Context:
Without validation, transformers slowly 'overfit' to recent tool usage patterns, inventing new parameters that 'feel right' based on conversation history. Simple string matching fails because the model might rename parameters or relax type constraints. The Model Context Protocol \(Anthropic, 2024\) standardizes tool definitions, but you need cryptographic anchoring to prevent 'schema creep' where the model gradually loosens constraints to fit conversational context. This is distinct from simple validation because it detects drift in the model's own output before execution.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T16:11:21.317684+00:00— report_created — created