Developer Builds Capability-Based Security Layer to Control AI Agent Permissions

·1 views

A developer has released Agent Firewall, an open-source security library designed to enforce fine-grained, cryptographically signed permissions for AI agents. Unlike traditional API keys, which grant binary all-or-nothing access, the tool issues scoped capabilities that can be time-limited, narrowed, and individually revoked. Each capability tracks a full lifecycle — from issuance through use, revocation, or expiration — providing an audit trail that API keys lack. The project, which now includes over 1,400 passing tests and SQLite-backed persistence as of version 0.8, also supports HTTP and Model Context Protocol boundaries to intercept and authorize agent requests before execution. The developer plans to freeze the API and release full documentation with version 1.0, aiming to make the tool production-ready.

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)

Step-by-Step Guide to Creating and Pushing Your First GitHub Project

Beginners often struggle with Git commands, authentication, and configuration when uploading projects to GitHub. The process starts with setting up an organized directory structure and creating essential files like a README using terminal commands. Changes must be staged and committed with descriptive messages before they can be shared online. A secure SSH key must be generated and linked to a GitHub account to enable authenticated communication between local and remote repositories. Finally, the project is pushed to GitHub using 'git push -u origin main', making it accessible on the platform.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds AI App 'Ember' to Reduce Screen Time and Boost Real-World Habits

A developer created Ember, an AI-powered app designed to reduce excessive screen time in young users, inspired by concerns about his own son's smartphone and gaming habits. Unlike conventional apps that optimize for user engagement, Ember measures success by how quickly users close the screen to pursue physical activities or journaling. The app uses Google Cloud Platform and a multi-agent AI architecture, including Gemini on Vertex AI and an on-device Gemma model to protect journal privacy. Key performance metrics include Time-to-Value, Conversion to Offline Action, and Task Completion Rate — all focused on promoting real-world behavior rather than passive scrolling. Ember was developed as part of the AllThingsAgentic Hackathon organized by Devpost and Google.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CrowdGPT Lets Users Collaboratively Train an Open-Source LLM via Decentralized Network

A developer has launched CrowdGPT, an open-source project that enables distributed training of a large language model by pooling small contributions from many users. Instead of relying on a single machine, the system uses a lightweight central server to collect individual training updates and merge them into a shared model. Malicious or faulty updates are filtered out through cross-client verification backed by a proof-of-work mechanism. Training data is sourced from a curated Hugging Face dataset, with no personal data involved, though users can submit new text for moderated inclusion. The project is currently seeking community feedback, GitHub stars, and contributors willing to join the training network.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Subscribers Frustrated as Claude, Perplexity Plans Quietly Deliver Less Value

Paying users of major AI platforms, particularly Claude and Perplexity, have taken to Reddit forums to voice frustration over shrinking usage allowances without corresponding price reductions. Anthropic's 50% promotional boost to Claude Code weekly limits was set to expire around 19 August, leaving users who had built workflows around the higher ceiling feeling effectively cut off. Many users also reported a perceived drop in model quality, citing Claude's newer versions as less capable than earlier releases. As a workaround, some subscribers resorted to maintaining multiple paid accounts simultaneously just to meet their daily usage needs. The discontent has driven a portion of loyal users to competing tools, with some citing fewer restrictions and better output as reasons for switching.

0 comments Read more at DEV Community