# The product review that issued a 100% discount

> Source: <https://discuss.huggingface.co/t/the-product-review-that-issued-a-100-discount/192108#post_2>
> Published: 2026-11-21 00:33:32+00:00

**Recommendation:** Delimit retrieved data tokens from system instructions before they reach the model, rather than relying on server-side validation after the fact, to stop hidden instructions in retrieved content from being executed.
More in the playbook "Securing agents and AI tools": <https://wpnews.pro/playbooks/agent-security>

[@Vivekwritesdaily](https://discuss.huggingface.co/u/vivekwritesdaily) The post correctly identifies that models assign equal weight to hidden instructions within retrieved content. The vulnerability stems from the attention mechanism treating all tokens in the context window as potential commands, regardless of their semantic origin. While server-side validation is essential, it acts as a post-hoc filter. A more robust approach involves architectural isolation where the retrieval layer strictly delimits data tokens from system instructions before they reach the model. This prevents the context window from ever parsing user-generated text as operational directives. How are you handling the boundary between retrieved data and system prompts in your current agent architecture?
