AWS Adds Domain and Date Filters to Bedrock AgentCore Web Search Tool

·1 views

AWS has introduced domain allow-listing and publish-date filtering for the Web Search tool within Amazon Bedrock AgentCore, its managed runtime for building AI agents. The new controls let developers restrict an agent's web searches to approved domains and exclude results older than a specified date, before those results reach the agent's reasoning process. This addresses a longstanding challenge in deploying AI agents, where models cannot independently assess source quality or freshness. Practical use cases include customer support agents limited to official documentation, research agents requiring up-to-date information, and compliance workflows that must audit which sources an AI was permitted to use. The filters also reduce exposure to prompt injection attacks by limiting reachable web content to a smaller set of known domains; pricing implications of the update have not been confirmed by AWS.

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)

Developer builds 129 browser-based tools with no backend infrastructure

A developer has launched Toolsholic, a collection of 129 web tools that operate entirely within the browser without any backend server. The site is hosted as a static export on a content delivery network, with all processing handled client-side. It uses libraries such as pdf-lib and pdf.js for document handling, the Canvas and File APIs for image tasks, and Web Workers for performance-heavy operations. The tools remain fully functional even when the device is offline, confirming that no server communication is required.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Architectural Firewall, Not System Prompts, Is the Real Fix for AI Prompt Injection

Developers at ThumbGate argue that relying on system prompts to prevent prompt injection in AI agents is fundamentally flawed, since language models cannot reliably distinguish instructions from data at the token level. The team has open-sourced an architectural approach that treats the LLM itself as untrusted and places a deterministic, code-based firewall between the model and any real-world tool execution. All external inputs are tagged as tainted and cannot alter tool invocation parameters, while every proposed action — such as sending an email or running a command — is intercepted and validated against a cryptographic allowlist in under one millisecond. The system has been deployed in production on a South Florida real estate deal desk, where thousands of public records are processed daily with no reported injection breaches. The developers say the solution can be run locally via a simple command-line tool, and urge builders to fix agent architecture rather than extending system prompts.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Background Jobs Silently Fail and How to Build Them Reliably

Backend engineer Maneshwar, creator of the LiveReview code-review tool, explains why offloading tasks like welcome emails to background job queues is essential for fast, resilient APIs. Running email sends inline during a signup request ties response times and uptime directly to third-party mail providers, which can be slow or unavailable. The recommended fix is to save the user, return a response immediately, and push a job to a queue such as SQS or RabbitMQ for a separate worker pool to process. However, this introduces a 'dual write' problem: if the application crashes after writing to the database but before enqueuing the job, the task disappears silently with no errors or alerts raised. This silent failure mode, where both systems appear healthy yet work is lost, is described as more dangerous than an outright crash because it goes undetected.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Drizzle ORM and TypeScript Can Optimize Backend Query Performance

Engineers at CSI TECH Informática have outlined best practices for building high-performance backend architectures using Drizzle ORM, TypeScript, and PostgreSQL. The article addresses a common bottleneck in enterprise applications: performance degradation caused by heavy ORMs that generate unnecessary queries, known as the N+1 problem. Drizzle ORM is presented as a lean, SQL-oriented alternative that gives developers direct control over query execution while preserving static type safety. The piece includes a practical TypeScript code example demonstrating an optimized join query between users and orders tables. The approach is positioned as suitable for microservices and corporate platforms requiring low latency, strict data integrity, and horizontal scalability.

0 comments Read more at DEV Community