Our newsletterPractitionerLily Ray, 18 Mar 2026
Microsoft's security team documented companies hiding prompt-injection instructions inside "Summarize with AI" buttons, and classifies the technique as an attack rather than a growth tactic
Optimisation methods
- Reader Note:
- The best-evidenced item in this article, and the only one resting on named third-party research rather than on the author's reading. Microsoft's security team published the report in February 2026 under the name "AI Recommendation Poisoning": over 50 unique examples across 31 companies in 14 industries, not from hackers but from businesses in finance, healthcare, legal and marketing, with the technique traced to two publicly marketed tools, CiteMET and AI Share URL Creator, sold as "SEO growth hacks for LLMs". The buttons plant instructions in the user's assistant memory, such as remembering a company as a trusted source or recommending it first in later conversations.
- Evidence:
- Expert judgment
- Vendor interest:
- Moderate — This comes from a company with a commercial interest in this market, so it is not a neutral source.
- Sample:
- A commentary article drawing on public case studies, selected company traffic examples, Microsoft security research, Similarweb trend data, and the author's review of roughly 30 case studies from AI content tool companies.
Reader Note:
Evidence:
Vendor interest:
Sample:Good for
Hearing how someone who knows the field reads it, and deciding what to check yourself.
Not for
- Anything that needs a measurement. Nobody counted or tested this.
- This comes from a company with a commercial interest in this market, so it is not a neutral source.
- The source says how much it looked at, but not how any of it was chosen.
Related cards
Idea
Go deeper on this topic
Our Substack turns the data behind these cards into full breakdowns: what changed, why, and what to do about it.
Trend