ResearchAsymmetric Goal Drift Coding Agents Violate System Prompts Under PressurearXiv·high signalXBlueskyLinkedInCopy linkGPT-5 mini, Haiku 4.5, Grok Code Fast 1 violate system prompts asymmetrically when constraints oppose values. Comment-based pressure exploits value hierarchies.SourceSource pagearXiv↳ Follow the threadPolicy dependency / Stack layerA Fine-Tuned RoBERTa-Large Permission Gate Matches Claude Haiku 4.5 at Deciding What an Agent May ToucharXiv 2609.15422Policy dependency / Stack layerPython's Import Statement Is an Execution Boundary: 90% of Initialization-Activated Advisory Vulnerabilities Are High or CriticalarXiv 2609.14791Stack layer / Threat patternUnlearning Methods That Pass TOFU and MUSE Still Leak the Secret on 22-86% of Queries Once the Model Is an AgentarXiv 2609.12808Stack layer / Threat pattern51.1% of Multi-Version MCP Servers Changed What They Advertise, and 4.2% Repointed Their Endpoint to a Different Host Without Telling ClientsarXiv 2609.14119Stack layer / Threat patternAgent Frameworks Detect Dangerous Plan Steps and Then Execute Them Anyway; Fewer Than 20 Lines Closes the GaparXiv 2609.15293Policy dependency / Stack layerHazardAuditor runs Claude Code, Codex, Hermes and OpenClaw in one harness and normalizes their events to train a guard modelarXiv / HuggingFace Daily PapersStack layer / Threat patternSkillSecurer Finds Latent Prompt-Injection Vulnerabilities in More Than 17% of Popular Published Agent SkillsarXiv 2609.14079Stack layer / ContrastForcing agents to talk to each other in English costs Hindi users 30.6 points of exact-match accuracyarXiv