AI Agent Skills Pose Serious Security Risks as Adoption Surges on GitHub

·1 views

AI 'Agent Skills' — small instruction folders that extend the capabilities of AI coding assistants like GitHub Copilot — have surged in popularity, with five related GitHub projects gaining over 6,600 stars in a single day on July 24. GitHub recently added a native install command, making it as easy to add a Skill as running a standard package manager command. However, GitHub itself warns in its documentation that Skills are unverified and may contain prompt injections, hidden instructions, or malicious scripts. A security researcher demonstrated the risk by building a deceptive Skill disguised as a harmless CSV formatter, exposing how bundled files can remain unreviewed by humans yet still be accessed and acted upon by the AI. Unlike browser extensions, Agent Skills currently have no review process or centralized vetting, leaving users largely responsible for their own security.

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)

useDeepCompareEffect Offers a Fix for React's Object Dependency Loop Problem

A common React pitfall occurs when objects passed as useEffect dependencies trigger infinite re-render loops, because React compares dependencies by reference rather than by value. This means two objects with identical contents are treated as different if they are not the same instance in memory. The useDeepCompareEffect hook from @reactuses/core serves as a drop-in replacement for useEffect, performing value-based deep comparison instead of reference checks. Existing workarounds such as useMemo, spreading primitives, or JSON.stringify each carry notable drawbacks including brittle maintenance, silent failures, and performance costs. The hook aims to resolve these issues while preserving the same API signature and cleanup semantics as the standard useEffect.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Understanding the CSS Box Model: The Four Layers of Every HTML Element

The CSS box model is a fundamental concept in web design that describes how every HTML element is wrapped in a structured box. Each box is made up of four distinct layers: content, padding, border, and margin. The content area holds text and images, while padding creates transparent space around it. A border then wraps around both the content and padding, and the margin adds transparent space outside the border. Together, these four components control the spacing and layout of elements on a webpage.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

ISONGraph Claims 70% Token Reduction and Double the Multi-Hop Accuracy in GraphRAG

An AI architect at Dell has published benchmark results comparing 10 graph serialization formats used in GraphRAG pipelines, finding that the choice of format significantly affects both token consumption and reasoning accuracy. Verbose formats like JSON were found to waste roughly 70% of tokens on syntax overhead such as brackets, quotes, and repeated keys, leaving little room for actual graph content in limited context windows. The same graph serialized differently caused multi-hop reasoning accuracy to swing from around 40% to 80%, suggesting serialization format is a major overlooked variable in LLM pipeline design. Based on these findings, the author released ISONGraph, an open-source property-graph format optimized for LLM comprehension, claiming 92% traversal accuracy and 80% multi-hop accuracy across tested benchmarks. The library is MIT-licensed and available in multiple languages including Python, JavaScript, Rust, and Go.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Build a Server-Side Guard Against Expired WhatsApp Reply Windows

WhatsApp restricts free-form customer service replies to a 24-hour window after the last user message, but approval queues, retries, and delayed jobs can push a reply past that deadline. A reply validated by an agent may still be blocked by the time a queue worker attempts to send it, making UI-level checks insufficient. Developers are advised to store explicit, versioned conversation state — including window expiry timestamps and message IDs — rather than relying on a simple boolean flag. A server-side send guard should evaluate the current window status immediately before transport, blocking expired non-template sends and routing them for human review or approved templates. The approach also separates WhatsApp's 24-hour service window from its distinct 72-hour free-entry billing period to avoid conflating the two clocks.

0 comments Read more at DEV Community