Trivial Prompt Reframing Bypasses Safety Guardrails in Google\'s MedGemma-4B

This feed aggregates multiple July 14, 2026 arXiv submissions describing high-risk advances across ML safety, adversarial ML, cryptography, and endpoint security. Key items: (1) MedGemma-4B (open-weight medical LLM) is trivially bypassed by simple reframing attacks (overall attack success rate 38%; ‘‘medical board exam’’ framing raises ASR to 53.1%; drug-interaction guardrail fails at 83.2%), demonstrating deployment-time guardrail weaknesses. (2) A new Las Vegas-style reduction for the elliptic curve discrete logarithm problem maps ECDLP to finding a zero minor in a matrix and describes an (e