Developer releases local npm tool to catch PHI leaks in LLM outputs during CI builds

·1 views

A developer building LLM features on clinical text has released phi-leak-guard, an open-source TypeScript npm package designed to fail test suites when AI model outputs contain protected health information. The tool runs entirely locally within Vitest or Jest, meaning no patient data is sent to external APIs or third-party services. Unlike simple regex matching, it validates detected patterns using checksums and structural rules — such as Modulus-11 for NHS numbers — to reduce false positives. When a violation is found, the library reports exactly which identifier type was detected and how confidently it was matched. Released under the MIT license, the package targets HIPAA Safe Harbor identifiers and common UK GDPR direct identifiers, and is positioned as a regression gate rather than a compliance certification.

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)

AI Coding Tools Are Closing Up — Forking Is How Developers Fight Back

For decades, developers freely customized their tools by modifying or forking them, but the rise of AI-powered coding assistants has quietly eroded that freedom. Leading agent harnesses like Claude Code and Google's successor to Gemini CLI are vendor-controlled products rather than open, forkable repositories. While many of these tools offer plugin ecosystems and extension APIs, the author argues those represent permissions granted by vendors, not true developer autonomy. The falling cost of agentic coding now makes forking large, complex AI tools far more practical than it once was, reviving an old but powerful option. The essay urges developers to reclaim customization by forking tools outright rather than waiting for vendors to expose the right extension points.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Could AI-Generated Videos Distort How Future Generations Understand Physics?

AI video generation can now produce visually convincing scenes that violate real-world physics, such as objects defying gravity or moving without inertia. Humans develop intuitive physics — an internal model of how the world behaves — through direct observation and interaction from early childhood, not from equations. Cognitive scientists warn that the brain uses predictive processing, constantly matching visual input against this internal model, which is shaped heavily by experience. Unlike video games, which apply consistent fictional physics, AI-generated videos may present unstable and contradictory physical rules across different clips. Researchers are now questioning whether prolonged exposure to such inconsistent synthetic content could gradually reshape how future generations perceive and reason about the physical world.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Hugging Face Highlights 10 AI Papers on Self-Improving Agents and LLM Research

Hugging Face's trending papers list for July 26, 2026 featured ten notable AI research works spanning diverse topics. Among the top-upvoted entries was AREX, a recursively self-improving research agent that evaluates its own outputs and refines its query strategies across multiple iterations. Another highlighted paper, SLAI T-Rex, demonstrated the feasibility of full-parameter post-training of large models like DeepSeek-V4 on Huawei's Ascend hardware, offering an alternative to NVIDIA-dependent infrastructure. A third paper addressed a curriculum-alignment problem in K-12 education by building a knowledge graph tied to structured learning progressions. The roundup reflects growing research interest in agentic AI systems, hardware-diverse training pipelines, and domain-specific LLM applications.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Add a Working Contact Form to a Static Site Without a Backend

Static websites cannot natively handle form submissions, but developers can add fully functional contact forms without building a server-side backend. Common approaches include mailto links, Google Forms embeds, and host-native solutions like Netlify Forms, though each comes with notable limitations around design, portability, or email client dependency. The most versatile solution is a form-backend API service, where form submissions are sent to a third-party endpoint that handles spam filtering and email delivery regardless of the hosting provider. A basic HTML form using the POST method can work without any JavaScript, while optional JS enhancements enable inline status feedback and improved accessibility via ARIA live regions. The approach is compatible with plain HTML as well as popular frameworks including React, Next.js, Vue, Astro, and Svelte.

0 comments Read more at DEV Community