Developer Builds LEASH System to Dynamically Restrict AI Agent Permissions in Real Time

·1 views

A developer created LEASH, an AI oversight gateway built in roughly one day for the Agents of SigNoz hackathon, designed to automatically revoke or restrict an AI agent's database permissions when errors spike. The system applies a tiered permission model — from full read-write access at T3 down to read-only at T1 — adjusting the agent's authority based on live telemetry rather than fixed rules. LEASH sits as a separate broker process between the agent and its tools, using SigNoz alerts and OpenTelemetry instrumentation to detect failure patterns within a five-minute window and trigger demotions via webhook. The architecture deliberately splits the agent and its policy enforcer into distinct services to prevent the agent from self-certifying its own trustworthiness. The approach adapts the SRE concept of error budgets — traditionally used to gate software deployments — and applies it to AI agent tool calls as a scalable alternative to manual human-in-the-loop review.

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)

Tool actlint audits MCP server safety labels, flags inconsistencies across 31 servers

A developer tool called actlint has been released to check whether safety annotations on MCP (Model Context Protocol) servers accurately reflect what their tools actually do. The tool scans a server's advertised tool list and compares each tool's name, description, and input schema against its declared safety labels, without ever executing the tools. When tested against 541 tools across 31 widely used MCP servers — including those from Stripe, GitHub, Microsoft, Google, MongoDB, and PayPal — only 11 servers came back fully clean. Four servers were found carrying labels that contradict capabilities their own schemas describe, a problem that can silently bypass safety confirmation steps in MCP clients. actlint can be run locally via a single npx command and supports CI integration, offline manifest replay, and baseline tracking to manage existing findings gradually.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How one invoicing app built five AI features without touching its legal records

Slate, an invoicing and bookkeeping SaaS built for Israeli businesses, has integrated five AI-powered features into its platform. The app uses Angular, Express, Prisma, and Postgres, and issues documents that serve as legally binding fiscal records under Israeli law. Israeli regulations require these records to be stored for seven years, numbered sequentially without gaps, and never altered after issue. These legal constraints directly determined how each AI feature was designed and scoped. The five features include receipt scanning, a books-based chat assistant, a support reply drafter, natural language document creation, and smart column mapping for Excel imports.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Safety Guardrails Found to Fail Consistently in Non-English Languages

Research covered by Dark Reading reveals that large language model safety guardrails, mostly trained on English-language data, fail to block jailbreak attempts when the same prompts are submitted in other languages such as French, Polish, or Finnish. The core issue is that refusal training and red-teaming efforts have historically concentrated on English attack patterns, leaving the safety layer undertested for other languages. No specific breach or victim has been reported; the finding highlights a structural vulnerability rather than a confirmed incident. This gap is especially significant for multilingual environments like the EU, where 24 official languages are in use, meaning attackers may need nothing more sophisticated than a translation tool. Semantic embedding-based approaches, which evaluate meaning rather than surface-level text patterns, are cited as a more language-agnostic defense compared to traditional keyword or regex filtering.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

JWT Authentication Explained: How Tokens Replace Passwords After Login

JWT (JSON Web Token) is a widely used authentication method in modern backend development that eliminates the need to repeatedly send passwords with every server request. When a user logs in with valid credentials, the server generates a JWT and sends it back to the client, which then includes it in all subsequent requests. The token consists of three parts — a header describing the algorithm, a payload carrying user details like ID and role, and a cryptographic signature that detects any tampering. Because the signature invalidates the token if it is altered, the server can trust its authenticity without maintaining server-side sessions. This stateless design makes JWT particularly well-suited for REST APIs and applications that need to scale across multiple servers.

0 comments Read more at DEV Community