FreoStream:Enhancing Stream Guardrails via Future-Aware Reasoning and Safety-Aligned Optimization

A multi-paper collection highlighting several high-risk issues and defenses across ML, software supply chains, hardware, and IoT. Key findings: FreoStream proposes a future-aware streaming guardrail to reduce over-refusal and improve jailbreak detection for token-level safety checks; SEVRA-BENCH demonstrates that LLM-based code-review agents can be socially engineered into approving PRs that reintroduce real-world vulnerabilities (CVE-linked fixes inverted), exposing a gap between closed- and open-source reviewers; Software Dark Matter (DARKFILES) finds large visibility gaps in SBOM-driven pip