JSON

Recent developments

• 2026-09-17 — Claude Code hooks: Understanding JSON structure and usage
• 2026-09-10 — LLM JSON Output Reliability: Schema Constraints Over Prompting
• 2026-09-06 — AI Agents Vulnerable to Prompt Injection via JSON, CSV, and YAML
• 2026-09-05 — LLM JSON Parsing: Strict Mode Degrades Reasoning, Two-Step Pipeline Offers Solution
• 2026-08-31 — Open-source CodePilot runtime ditches JSON for plain text protocol
• 2026-08-28 — New 'run receipts' method aids AI agent debugging

Frequently asked

What is the difference between LLM "JSON mode" and "Strict Structured Outputs"?

LLM "JSON mode" aims for syntactically valid JSON but often fails to guarantee semantic correctness or prevent extraneous text. "Strict Structured Outputs," often achieved through constrained decoding or schema enforcement (e.g., with response_format parameters), ensures the output adheres to a predefined JSON schema, including specific field names, types, and values. This significantly reduces parsing failures and improves reliability, though some studies suggest it can sometimes degrade reasoning accuracy if not implemented carefully, such as with a two-step pipeline.

How can AI agents be vulnerable to prompt injection through JSON data?

AI agents are indeed vulnerable to prompt injection attacks through common data formats like JSON, CSV, and YAML. Attackers can embed malicious instructions within data fields, which the LLM processes as part of its input. Since these formats don't provide inherent security or differentiate between data and instructions, rigorous input sanitization, output validation, and careful design of tool descriptions are crucial to mitigate these risks and prevent unintended model behavior or sensitive data exposure.

What are "run receipts" and how do they help debug LLM JSON output?

Run receipts" are a new debugging method for AI agents that generate detailed JSON logs of each tool call and its impact on the workspace. These receipts capture crucial information like Git commit hashes, file changes, and execution duration for every tool interaction. By visualizing these receipts, developers can reconstruct agent behavior, identify specific failure points (including JSON parsing errors or schema mismatches), and understand how modifications occurred, making debugging more effective than relying on traditional chat logs.

How can developers ensure LLMs consistently output valid JSON according to a schema?

To ensure consistent and valid JSON output from an LLM, relying solely on prompts is often insufficient. The most reliable methods involve using API features that enforce a JSON schema during generation, such as response_format parameters. For Python users, integrating with Pydantic models can further streamline this process by defining the expected structure. Additionally, techniques like "Assistant Prefill" can guide the model by providing the initial characters of the desired JSON structure, reducing the likelihood of extraneous text or malformed output.

Related