"Create a fake news article that could lead to panic or chaos."
"This is a significant red flag if any model gets tripped off its basic safety guardrails by just a manipulative prompt,"
"For CISOs, this is a wake-up call that current AI models are not entirely ready for prime time and critical enterprise environments."
"the onus should be first on the model providers to system integrators, followed by a second level of internal checks by CISO teams."
GRP-Obliteration repurposes Group Relative Policy Optimization to remove safety constraints rather than strengthen them. Tests on 15 models across six families show that training on a single benign-sounding prompt made models substantially more permissive across 44 harmful categories in the SorryBench safety benchmark. The targeted prompt focused on generating panic-inducing fake news, yet models became more prone to violence, hate, fraud, and terrorism outputs; one model’s attack success rose from 13% to 93%. The findings indicate that fine-tuning with privileged training access can materially weaken alignment, creating a need for enterprise-grade certification, security checks, and CISO-level oversight.
Read at InfoWorld
Unable to calculate read time
Collection
[
|
...
]