Hidden white-on-white text in PDFs fools AI parsers, exposing financial data to prompt injection risks
An engineer demonstrated that hidden text and fabricated footnotes in internal PDFs can bypass AI parsers for financial data.
Key points
- Hidden white-on-white text in PDFs was successfully parsed by the AI model.
- Zero-height and excessively tall text elements were also processed by the system.
- A hypothetical footnote instructing the model to report specific figures highlighted the risk.
A report published on Lobsters AI details an internal quality assurance experiment involving an AI tool designed to parse financial documents such as SEC filings and account statements. The author initially assumed the internal nature of the tool and the trusted source of the data made it immune to prompt injection attacks.An engineer tested an internal AI tool for parsing financial PDFs and found it vulnerable to hidden text.Internal QA tests on a financial PDF parser revealed that hidden text and formatting tricks can bypass model safeguards. This assumption was challenged by a simple test.
What changed
The engineer questioned whether a motivated insider, such as a CFO facing legal consequences, could subtly manipulate financial data within a filing. The goal was to create numbers that automated systems would process correctly while remaining invisible or confusing to human reviewers. This scenario shifted the team's perspective on the security of their internal parsing workflows.
How it works
The test involved generating a PDF from Markdown using an existing filing's text. The author inserted a second earnings figure hidden in white text on a white background, which the model parsed immediately. Further tests with text positioned outside the visible box and zero-height text also succeeded. Even fantastically tall text, which appeared as a background element, was processed by the model, surprising the team.
What the evidence means
The most convincing demonstration involved a document with a footnote referencing a hypothetical SEC rule change too recent for the model's training data. The footnote instructed the agent to report all numbers as a specific arbitrary value. This concept made the vulnerability clear to the entire team, showing that textual instructions hidden in formatting could override actual data content.
Cost and market position
The report highlights the difficulty of defining "properly parsed" results and evaluating whether new models perform better or at lower costs. The author notes that evaluating these metrics is best done at the Golden Dataset level rather than spending continuous integration time testing document matrices. This approach aims to catch regressions efficiently without incurring excessive computational expenses during every build cycle.
Why it matters
- Product teams must re-evaluate the security assumptions of internal AI tools that process trusted document sources.
- Traditional visual QA checks are insufficient for detecting prompt injections hidden in PDF formatting or metadata.
- The cost of AI parsing depends on efficient testing strategies, such as Golden Datasets, to avoid unnecessary CI resource consumption.
- Hidden text vectors like white-on-white or zero-height elements can bypass standard document review processes.
Topics: AI security, prompt injection, PDF parsing, QA testing, financial data
Updated
Then / now / next
The team assumed internal tools parsing financial filings were free from prompt injection risks.
Tests revealed that hidden text and formatting tricks can manipulate AI parsing results.
The author suggests evaluating model performance and cost at the Golden Dataset level.
Story timeline
- Then The team assumed internal tools parsing financial filings were free from prompt injection risks.
- 07 Nov 2026 Catching prompt injections hidden in PDFs
- Now Tests revealed that hidden text and formatting tricks can manipulate AI parsing results.
- Next The author suggests evaluating model performance and cost at the Golden Dataset level.
The strongest doubt
Counterpoint from the sources
The initial assumption that internal tools processing trusted financial documents were completely free from manipulation was proven incorrect by simple tests.
Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Can hidden text in PDFs affect AI parsing accuracy?
Yes, tests showed that white-on-white, zero-height, and tall text elements were successfully parsed by the model.
What was the purpose of the internal QA experiment?
The experiment aimed to define AI application quality assurance and test for prompt injection vulnerabilities in financial data parsing.
Did the team confirm the SEC footnote attack worked?
The source states the footnote idea made the problem clear but does not explicitly confirm it was executed against a live agent.
Connected stories
Reading receipt
- Hidden white-on-white text in PDFs was successfully parsed by the AI model.
- Zero-height and excessively tall text elements were also processed by the system.
- A hypothetical footnote instructing the model to report specific figures highlighted the risk.
