Hidden text in PDFs could steer AI tools reading bank statements and invoices

·4 views

A PDF can contain a hidden text layer that people never see but software reads, which becomes a risk when language models process documents. Redact reported in October that WhatsApp help pages carried hidden instructions aimed at an AI, found on 80 of 545 pages, which WhatsApp attributed to a bug and said targeted Meta's own chatbot. Financial documents are a concern because users increasingly ask models to read statements and invoices, and a researcher at Snyk showed in 2025 that white text in a statement PDF could shift a model's credit rating from poor to excellent. The author, who builds a template-based bank statement extractor with no model in its extraction path, tested whether hidden content could affect their own system by planting it in copies of a test statement. The writeup notes no public production case has been found so far.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Several Linux kernel vulnerabilities accidentally posted to public mailing list

Details of several Linux kernel security flaws were inadvertently published to the public Linux Kernel Mailing List, according to a Phoronix report. The disclosure appears to have occurred before the vulnerabilities were intended to become public, raising concerns about exposure. The article linked from the Hacker News post covers the incident and the broader trend it reflects. The available source material is limited to the link and metadata, so further specifics about affected versions or patches are not confirmed here.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Offline-First App Uses On-Device AI to Detect Crop Diseases in Central America

Smallholder farmers in Costa Rica and Central America often work in areas with little or no cellular coverage, making cloud-based disease diagnosis impractical. To address this, developers built HealthyCropping, an open-source, non-commercial Android app that runs its core functions, including photo capture, inference, and severity assessment, without an internet connection. The app uses specialist MobileNetV3 models for coffee leaves, cacao pods, and banana leaves, quantized to under 10 MB each and executed through native TFLite runtimes in under 50 milliseconds. To prevent accuracy loss from differences between Python training and mobile preprocessing, the team built a shared fixture test suite that checks for numerical parity between the two environments. Results are stored locally and synchronized to a cloud backend once connectivity returns.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Is it a button or a div? A quick check for keyboard-accessible clickable elements

Landing pages often contain elements that look like buttons, such as menu toggles and tabs, but are built as generic tags with click handlers. Such elements are not reachable by keyboard or announced correctly by screen readers, unlike native button elements. The W3C's WCAG 4.1.2 guidance requires that custom controls expose the correct name and role to assistive technology. The author also notes that a button inside a form submits by default unless type="button" is set. A simple test is to tab through the page, check each unreachable element's tag, and verify that Enter and Space activate it.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

PHP Patches Two Critical Vulnerabilities as Java, Node.js, and Railway Also Face Security Issues

On October 21, PHP released fixes for two critical remote code execution flaws, CVE-2026-109618 and CVE-2026-109619, which carry a CVSS v4.0 score of 9.2. The same day, Oracle issued an out-of-band alert for a Java vulnerability, Node.js closed ten flaws, and Railway disclosed that customer secrets had been accessed. A day earlier, a version of the OVHcloud command-line tool was found to have shipped with malware, and affected users are advised to rotate their secrets. According to the PHP Foundation, the underlying problem was first noticed in 2024 but stalled over disagreements about competing fixes before two research teams independently found a working attack path this summer. Security teams are urged to update PHP to the patched releases and restart their services, while no confirmed successful compromises have been reported so far.

0 comments Read more at DEV Community