Stack layer / Update thread
SLBench Finds Up to 70% Unsafe Execution When Codex and Claude Code Follow Logical Relations in Skill Files
arXiv
Policy dependency / Contrast
Make context compaction a trained policy, not an inference-time heuristic — CompactionRL lifts SWE-bench up to +7 points
arXiv 2607.05378
Policy dependency / Stack layer
Reason Less, Verify More: Deterministic Gates Catch a Silent Policy-Violation Failure in Tool-Using Agents
arXiv
Contrast
SMetric Rethinks LLM Scheduling for Agentic Serving, Where KV-Cache Reuse Tops 80%
arXiv
Stack layer / Threat pattern
Mitigating Taint-Style MCP Server Vulnerabilities Without Touching Server Code (SPELLSMITH)
arXiv
Contrast
Shared Selective Persistent Memory Lifts Agentic Task Completion to 96% and Cuts Per-Call Token Cost 97x
arXiv
Policy dependency / Stack layer
HalluSquatting weaponizes LLM hallucinations to push botnet malware into agents
The Hacker News
Stack layer / Threat pattern
GitHub Ships CodeQL 2.26.0 With AI Prompt-Injection Detection for JavaScript/TypeScript
GitHub (CodeQL release notes)