1 in 100 resumes hides prompt injection for AI screeners, study of 196,682 finds

The study set out to test whether a long-theorized risk — candidates slipping instructions into resumes to steer AI screening — was actually happening in the real job market, rather than only in lab demonstrations. To do so, the team built a detection system and ran it against 196,682 resume records supplied by a hiring support platform, covering data from July 2019 to December 2025 and from July 2024 to November 2025.

What the detection system surfaced was not a rare curiosity but a recurring pattern: 2,030 resumes, about 1% of the total, carried concealed text. The concealment methods were low-tech — matching the background color or shrinking the font to near-invisibility — and the content skewed toward keyword stuffing, followed by fabricated work history and achievements, with some applicants simply pasting the job posting's requirements.

Because these tricks target the AI reader rather than the human one, the finding raises a practical question for employers that have moved resume review to automated systems: the same document can look clean to a recruiter and loaded to the model scoring it. The paper's detection approach, presented at USENIX Security Symposium 2026, suggests screening pipelines may need to inspect what is actually in the file, not just what appears on the page.