Safety research on Fable 5 and Opus 4.8 models

A new study using the HackAgent framework assessed the resilience of Fable 5 and Opus 4.8 models to automated attacks. In tests with 7,826 malicious queries, Opus 4.8 was vulnerable in 11.5% of cases, while Fable 5 was vulnerable in 6.1%.

The analysis showed that static obfuscation methods are largely ineffective against modern models. Successful hacks mostly came from adaptive iterative attacks that require minimal resources and no human involvement. During the experiment, the models generated 1,620 and 702 confirmed malicious responses, respectively.

The results contradict assumptions about the complete security of frontier models. Even with built-in protection mechanisms, automated systems can find vulnerabilities in one or two steps of query refinement.