Meta AI Updates: October 31, 2026

1. Meta Moves LLM-Generated “Catching” Just-in-Time Tests Into Production Across Its Codebase

Meta published details on Catching Just-in-Time Tests (JiTTests), an LLM-based system that generates executable tests to expose bugs in code changes before they land, which Meta says has moved from research prototype to part of its production infrastructure. Unlike LLM code review, which judges a diff the way a human reviewer would, a JiTTest demonstrates a concrete behavioral difference the change introduces. Two workflows run in tandem: an intent-aware workflow infers the diff’s intent and risks, builds mutants representing plausible bugs, and generates tests to kill them, while a “dodgy diff” workflow treats the diff itself as a mutant and generates tests that distinguish its behavior from its parent. Overnight, on spare capacity, the system targets the previous day’s risky diffs using a method similar to Meta’s Diff Risk Score, then passes weak catches through an LLM-as-Judge ensemble and a deterministic rule-based filter called RubFake before ranking them and sending the top candidates to engineers. Meta reports that 589 engineers have engaged with the results (64% of outreach), that engineers report a 60% true-positive rate, and that in the first half of 2026 the system helped catch 74 serious privacy bugs at the diff stage while privacy SEV occurrence fell 20%. The “human-on-the-loop” design means engineers set criteria and confirm whether a flagged behavioral change was intended rather than reading test code; Meta notes the approach costs more compute than LLM review and that LLM judgments are non-deterministic, which is why the rule-based filter exists. Source