ResearchSlowBA: Efficiency Backdoor Attacks on GUI AgentsarXiv·high signalXBlueskyLinkedInCopy linkNovel attack targeting agent efficiency not correctness. Triggers cause excessive reasoning steps. Agents work correctly but become unusably slow.SourceSource pagearXiv↳ Follow the threadStack layer / Threat patternContext Privilege Escalation Attacks Hit 12 Real Agent Harnesses, Including Claude Code and CodexarXiv 2609.01222Policy dependency / Threat patternCORAL Puts an LLM Agent in a Live Closed Loop on Production Recommenders and Wins Two A/B TestsarXiv 2609.02730Policy dependency / Stack layerCo-evolving the harness alongside the policy cut AgentDojo attack success 3x while raising benign utilityarXiv 2609.02786Policy dependency / Stack layerTyped Provenance Guardrails Block All 19 Unsafe Releases From a Persistent Agent's Autobiographical MemoryarXiv 2609.02127Stack layer / Threat patternmzCache Cuts On-Device LLM Time-to-First-Token 2.1-5.5x When the OS Evicts Model MemoryarXiv 2609.01338Stack layer / Threat patternSCX Router Scores Model Suitability With No Autoregressive Generation and a Persistent Text-Only KV CachearXiv 2609.02292Threat pattern / ContrastA case study watched one coding agent build a real data system against a fixed spec, and catalogued what it got wrongarXivStack layer / Threat patternTwo Ways to Make Alignment Evaluations Look Like Real Deployments, and They ComposearXiv 2609.02302