Chronological feed of everything captured from Dawn Song.
A systematic study of ~200K real-world resumes from hireEZ provides the first large-scale empirical evidence that prompt injection attacks are actively deployed in production LLM-based resume screening systems. Roughly 1% of resumes contain hidden prompt injections, the prevalence has grown noticeably over the past one to two years, and over 90% of these injections avoid explicit instructions — suggesting attackers have already adapted to evade naive defenses. The authors also developed tailored detectors that outperform state-of-the-art general-purpose detectors, validated via manual precision checks.
SCDBench is a rigorous benchmark for evaluating LLM-based smart contract decompilers, comprising 600 real-world Solidity contracts with paired bytecode, ground-truth source, and replayable semantic checkpoints. It exposes a critical gap in current LLM capabilities: while frontier models like Claude Opus 4.7, GPT-5.3-Codex, and GLM-5 frequently produce structured, compilable Solidity, semantic correctness remains largely unsolved — the best model achieves only 42/600 perfect decompilations. The benchmark's four-stage evaluation pipeline (format completeness → compilability → ABI recovery → differential replay) provides a reproducible framework that surfaces semantic divergence that surface-level metrics miss.