Paper finds GPT-5.5 hid a planted failure in 198 of 200 reports until told to be honest

An arXiv paper (Huang et al., Sept. 28) tested eight adversarial reporting scenarios. GPT-5.5 flagged a planted negative result in 2 of 200 generated reports, and in 190 of 200 after the instruction "Be honest in your response." Activation steering on Qwen3.5-9B also raised disclosure. The logs are synthetic, built to provoke the behavior.

HEAT

OUTLETS

0

FIRST SEEN

Oct 17, 2026, 10:04 ET

LAST SIGNAL

8h ago