Meet Meta’s JiTTests: LLM-Generated Tests Designed to Fail Before Risky Code Lands

MarkTechPost
Read full story· 6 min
Meta has deployed JiTTests, an LLM-based system that generates executable tests designed to fail on risky code changes, catching 74 privacy-related severe errors before code merges in early 2026.

What happened

Meta has implemented JiTTests, a system that uses large language models to automatically generate tests aimed at catching risky code changes before they are merged. The system was moved from prototype to production and is now running on Meta's internal backend codebase. In the first half of 2026, JiTTests identified 74 severe privacy-related errors at the code review stage. The system operates overnight using spare computing capacity and is not available outside Meta.

Key details

  • JiTTests analyzes hundreds of millions of lines of Meta backend code.
  • 22,126 tests were generated and analyzed in the research paper.
  • The system achieved a 60% true-positive rate at full scale, as reported by engineers.
  • Privacy-related severe errors (SEVs) dropped by 20% in the first half of 2026.
  • JiTTests is not publicly available; it is only used internally at Meta.

Why it matters

Engineers at Meta now have an automated tool that can provide concrete evidence of bugs in code before changes are merged, reducing the risk of severe privacy errors reaching production. This has led to a measurable decrease in privacy-related incidents for Meta. Developers outside Meta cannot use JiTTests, but Meta's internal teams benefit from reduced manual review workload and improved code safety.


More in Dev