FAVA: Formal Authorization for Verified Agents with Evidence-Backed Permission Graphs

This collection contains recent research on agent authorization, autonomous cyber-defense benchmarking, LLM jailbreaks, cryptographic performance and verification, E2EE transcript-consistency attacks, automated threat modeling, and diffusion watermark robustness. The most directly security-relevant findings are RoguePrompt's high reported rates for moderation bypass and instruction reconstruction, and Send and Pretend's demonstration that major E2EE group messengers may permit transcript equivocation, selective content manipulation, and poll rigging. These are research findings rather than CVE