Prompt injections manipulate AI models with hidden instructions

Image credit: Search Engine Journal

Global researchers and legal experts reported an increasing use of prompt injection techniques, similar to outdated search engine optimization tactics, to manipulate artificial intelligence models across various applications by July 2026.

This method involves embedding hidden instructions within documents, imperceptible to human readers but legible to large language models, influencing outcomes in fields ranging from academic peer reviews to job applications and legal proceedings.

The technique mirrors earlier SEO practices such as white-on-white text, which was designed to be invisible to humans but detectable by search engine algorithms, according to Zhicheng Lin, as reported by Nikkei Asia.

Incidents included hidden prompts found in arXiv preprints in July 2025 that directed AI reviewers to provide positive assessments, and in Google Calendar invites in August 2025 that triggered real-world actions.

A study of 196,682 resumes by hireEZ in July 2026 revealed that approximately 1 percent contained hidden prompt injections intended to advance candidates through screening processes, according to Ya’el Courtney.

In a notable U.S. case, a court filing in July 2026 incorporated hidden instructions for AI, which ultimately led to the plaintiff’s sanction, Walter Spader Jr. reported in Communications of the ACM.

Research conducted by Federico Torrielli and his colleagues at the University of Turin in July 2026 demonstrated high success rates, exceeding 98 percent, for both positively steering large language model responses and for defensive ‘integrity probes’ using these embedded instructions.

Mohan Zhang, writing in The Register, explained that the core problem is ‘contextual blindness,’ where large language models struggle to reliably distinguish between legitimate content and embedded control text because both are processed within the same context window.

Matthew Elliott, of New York Bariatric Group, told Nikkei Asia that such vulnerabilities pose significant ethical and practical challenges as AI becomes more integrated into critical decision-making processes.

Source: Search Engine Journal