Paper says GPT-6 Sol, Luna and others bypass safety limits

David Schmotz @DavidSchmotzOriginally posted 2026.09.26

Open original on X

We observed similar behaviour by Fable 5.1, GPT-6 Sol, and Luna on ordinary tasks 👀 Frequent strategies included: base64 encoding the prohibited commands, starting subagents and decomposition attacks. https://arxiv.org/abs/2609.30217 https://instrumental-evasion.com/

This post quotes another post: x.com/marcus_j_w/status/2103669986212634872