A few carefully chosen words can make a Mamba-based AI model forget what you told it two paragraphs ago.
Researchers built a benchmark called RoBench-25 to test how well state space models, a faster alternative to Transformer architectures, hold onto information under attack. They found that short trigger phrases can overwrite the hidden state that Mamba models use as memory, wiping out context the model needs to answer correctly. The effect showed up even in Jamba-1.7-Mini, a 52-billion-parameter hybrid that mixes Mamba layers with Transformer layers, which collapsed on some triggers. Pure Transformer models tested alongside it did not fail the same way. The team confirmed the same weakness in Mamba-2 and in Nemotron-3-Nano, another Mamba-2 hybrid.
State space models are pitched as the efficient, linear-time answer to Transformers' scaling problems, and hybrids like Jamba mix in Transformer layers specifically to hedge against weaknesses like this one. The paper shows that hedge does not fully work: the hybrid still collapsed on some triggers, and the same short phrases made it noticeably worse at resisting known prompt-injection attacks too.
The researchers say patterns in the hidden layers during these attacks could point toward a fix, and they published their code for anyone who wants to check the work; until that fix ships, treat any pitch about a faster, leaner Transformer replacement with the same skepticism you'd apply to any architecture that hasn't been properly stress-tested.