Talking an AI out of its own safety rules.
Korpalis selects material that helps builders and business teams understand what changed, why it matters and what to check next. Every entry below includes an original explanation and a direct link to its publisher.
Researchers demonstrate that even benign fine-tuning can severely degrade LLM safety alignment by selectively sharpening pathways in model internals that bypass refusal mechanisms, revealing a previously underexamined geometric explanation for alignment fragility.
Researchers demonstrate that jailbreak robustness varies significantly across different operational states—even system prompts unrelated to safety can alter attack success rates. This finding challenges the assumption that single-configuration robustness measurements reflect real-world resilience.