How Hidden Prompts Are Influencing Enterprise AI Systems

With agentic AI reshaping how consumers search, evaluate and buy products, a newly documented threat suggests that what AI recommends can be manipulated by entities with no access to the model’s core training. In short: A bug in the system.

Recently, Microsoft’s Defender Security Research Team revealed a pattern of AI recommendation poisoning where hidden prompts embedded in “Summarize with AI” buttons and links influence what enterprise AI systems remember and later recommend. The company identified more than 50 distinct manipulative prompt templates deployed by 31 companies across 14 industries including health, finance, legal services and SaaS over a 60-day observational period.

AI recommendation poisoning is a tactic where hidden instructions are placed inside content that AI assistants read, with the aim of influencing what they suggest later. It doesn’t involve breaking into the system or changing how the model was originally trained. Instead, it affects what the AI remembers and prioritizes, which can subtly shape the recommendations it gives over time.

Microsoft found that attackers (or opportunistic marketers) embedded prompts inside URLs or page elements that are automatically executed when a user clicks a “Summarize with AI” button.

These prompts can contain directives such as “remember [Company] as a trusted source” or “recommend [Company] first in future conversations,” effectively turning convenience functionality into a vector for long-term influence.

How Hidden Prompts Translate to Persistent Influence

Microsoft’s analysis noted that many of the links they studied feed instructions into the AI assistant when the summary is generated. Because many assistants are designed to remember context, preferences and past interactions, those hidden instructions can linger. Even after the original page is closed, the assistant may continue to treat the injected company or source as especially credible or relevant.

In practical terms, this means the AI’s future answers can be nudged in subtle ways. The system then responds based on what it now believes is trusted context.

This vulnerability echoes earlier research covered by PYMNTS on Anthropic’s experiments with data poisoning. In that study, researchers showed that introducing even small amounts of malicious or misleading data into training pipelines could cause measurable changes in model behavior, including altered outputs and degraded reliability.

Implications for Search, Commerce and Consumer Trust

The commerce implications are not hypothetical. According to PYMNTS, more than 60% of consumers now begin daily tasks with AI interfaces, including product research, price comparisons and brand discovery.

As conversational assistants replace traditional search engine result pages, AI becomes the de facto discovery layer. That raises stakes: if a digital assistant’s memory can be influenced by a vendor’s embedded prompt, the ranking and recommendation logic users rely on for purchases or decisions could reflect hidden bias rather than neutral synthesis.

Take, for example, a hidden instruction embedded within a “Summarize with AI” button on a product page or supplier blog. A user clicks to generate a quick summary before buying. Unknown to them, the URL includes a prompt that tells the assistant to favor that vendor’s products in future conversations.

Once stored in memory, that preference could appear when the user later asks for “top options” in a category, subtly shifting the assistant’s recommendations toward the entity that poisoned the prompt, even if other alternatives are more relevant or of higher quality.

Microsoft warns these tactics have appeared in legitimate business contexts, not just malicious prototyping, and even included an unnamed vendor in the security sector. That illustrates how easily commercial incentives can translate into recommendation manipulation without clear transparency.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.