Risk Models as Mediating Artifacts: A Postphenomenological Analysis of the CIIM Framework in Cybersecurity Practice

A batch of recent arXiv papers surveying and advancing security-relevant research across multiple layers of ML and systems stacks. Key highlights: AutoRISE and other automated red‑teaming work move from prompt search to program/strategy search, substantially increasing jailbreak/attack success; RouteGuard demonstrates that skill‑poisoning of LLM agents produces an internal ‘attention‑hijack’ signature and that internal‑signal detectors outperform lexical filters (F1≈0.8834, recovers ≈90.5% of attacks missed by text screening); the Reconstructive Authority Model (RAM) argues attestation alone (