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

arXiv 2605.12529•37ad203f0d64821522920c4755df14b92512812d3e93f9c0b038795a7ee7766f
CPythonCoT-monitoringLLM-backdoorsLuaQuickJSRoPESafeContextautonomous-drivingbackdoor-detectioncontext-assemblydecision-time-assemblyhard-brakinghidden-objectivesjailbreaksmodel-extraction-benchmarks','GNN-extraction','watermarking-defemodel-unlearningoverride-hookspersona-conditioned-attacksphysical-adversarial-camouflagered-teamingscript-runtimessemantic-fuzzingsmall-model-monitoringtrajectory-manipulationwatermark-preservation

Paper metadata

arXiv ID
2605.12529
Version
Not specified by this published record
Category
Computer Science — Cryptography and Security (cs.CR)

The PDF link points to arxiv.org. Baitaphish does not expose a private stored PDF.

Evidence and limitations

Source ID
arxiv_cs_cr
Record identifier
37ad203f0d64821522920c4755df14b92512812d3e93f9c0b038795a7ee7766f
Enrichment time
2026-05-14T07:23:35Z
AI-assisted enrichment
Yes

This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.