Catching split attacks across compactions in long-context coding agents

Researchers analyze vulnerabilities known as split attacks in long-context coding agents, focusing on how these attacks persist through compaction processes in AI models.