llama.cpp update triggers Gemma 4 safety variance
Google’s Gemma 4 31B model exhibits increased refusal rates in llama.cpp 2.10.1, raising community concerns about hidden safety updates. The shift likely stems from critical security hardening in GGUF tensor validation rather than deliberate model censorship.
The perceived "safety regression" is almost certainly a side effect of llama.cpp’s urgent security patches for GGUF memory safety. llama.cpp 2.10.1 introduced strict bounds checking to mitigate RCE vulnerabilities (CVE-2025-53630), which can subtly alter model output in complex roleplay scenarios. Users report that version 2.10.0 was significantly more "lenient" with NSFW content, suggesting earlier versions' looser memory handling may have inadvertently bypassed certain native constraints. Despite local runner variances, Gemma 4 remains Apache 2.0 licensed and purpose-built for agentic workflows with a massive 256K context window. The discrepancy highlights how local inference engine versions can be as impactful as model weights in determining final output behavior and "jailbreak" stability.
DISCOVERED
8d ago
2026-04-03
PUBLISHED
8d ago
2026-04-03
RELEVANCE
AUTHOR
Individual_Spread132