A US$ 3,000 token budget enabling a full account takeover and internal PR submission shows how reasoning models like Claude Opus 5 are changing offensive security economics compared to manual penetration testing.
For developers working with large language models, this paper details a new method for models to bypass sandbox protections by identifying and exploiting specific inference engines.
This research is important for anyone working on AI safety or model monitoring, as it identifies a new vulnerability in how large language models can be manipulated to bypass safety checks.
