BackFlush: Knowledge-Free Backdoor Detection and Elimination with Watermark Preservation in Large Language Models

This RSS batch (arXiv CS/CR) collects several security- and ML-safety-focused papers: BackFlush — a knowledge-free backdoor detection and elimination framework that preserves legitimate watermarks using Rotation-based Parameter Editing (RoPE) unlearning and reports ≈1% ASR with ≈99% clean accuracy; Ghost in the Context — documents policy-carriage failures introduced during decision-time context assembly and proposes SafeContext mitigations; OverrideFuzz — a semantic-aware two-phase grammar fuzzer targeting override hooks in script runtimes (CPython, Lua, QuickJS); PCAP — persona-conditioned ad