Sep 5, 2026

Reverse engineering stood out as the clearest theme. Jane Street's 2026 puzzle shipped only a GDSII layout file for a chip that checks an 11x11 Star Battle puzzle; solvers had to reverse the physical mask format back into a verifiable state machine, matching standard-cell footprints against the sky130 PDK and rebuilding a full netlist from metal and via tracing. Separately, a May 2026 project produced a standalone Android client for Meta Ray-Ban Stories by reverse engineering the firmware, allowing users to download full-resolution media and manage the device without the official Meta app, with the Meta AI app needed only once to extract secrets.

AI evaluation and agent governance faced a reliability reckoning. A preregistered study found language model judges, which increasingly gate training data and leaderboards, showed significant repeatability discrepancies across 52,988 attempts, challenging their stability as measurement instruments. In parallel, an engineer documented 14 months of refining rules for Claude Code, describing a demotion ladder that highlighted machine limitations; another project used parallel AI agents to keep a London Underground app consistent across four repositories, 458 commits and 42 releases, while noting the costs.

Platform and tooling improvements emerged for high-volume and local development. C# 15 previews non-boxing union types for .NET 11, reducing GC pressure in high-volume scenarios, while Yulin offers a local AWS simulator for Node.js applications that runs tests and applications in-process without external network calls. A separate analysis tested whether a business's operating rules could be recovered from its data; only six rules were recovered, but eighteen previously undocumented constraints emerged.

Security and soundness gaps drew attention. A pentester's guide to Salesforce LWR argues Experience Cloud guest users have richer querying through WebRuntime, UI-API and GraphQL endpoints, making misconfigurations a serious vulnerability. On the compiler side, an empirical study of 30 reported GitHub issues found rustc contains soundness bugs that allow programs to pass type checks that should fail, undermining Rust's type-sound reputation.