Topowatch tool measures how workspace structure affects AI agent prompt injection risk

·3 views

Security researcher Pedro Sordo Martínez has released topowatch, an open-source command-line tool that audits how a code workspace's topology influences its vulnerability to indirect prompt injection attacks. The tool is based on research published in arXiv:2608.14876, which found that workspace structure measurably affects the Attack Success Rate (ASR) of malicious payloads targeting agentic coding assistants. In benchmark tests using 200 trials with a fixed seed, flat monolithic workspaces showed a 100% ASR, while modular and deeply nested layouts dropped to 0% due to the agent reading less of the workspace. The current v0.1 release uses a synthetic agent rather than a real coding assistant, with future versions planned to test against tools like Claude Code in a sandboxed environment. The project is available on GitHub under the AGPL-3.0-or-later license and can be installed via pip without requiring API keys or network access.

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 Times Every Hour Building a Field Service Management System From Scratch

A developer and maintenance manager built a custom field service management system called FieldOS to accurately determine how long such software takes to build, aiming to answer client pricing questions with real data instead of estimates. The system covers five core functions: job request tracking, parts inventory management, technician dispatching, and automated invoicing integrated with an established payment processor. During integration testing, a low-stock alert silently stopped triggering when the parts and job-tracking modules were connected, traced to a startup sequencing error where one module checked the other's readiness before it had fully initialized. A separate invoicing bug produced correct totals but blank line-item descriptions, rendering the invoice unusable for customers despite the accurate final figure. The project is being documented as a multi-part series, with this first installment covering the core system build and the real engineering challenges that only surfaced when all components ran together.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Nvidia Takes Minority Stake in Power Infrastructure Firm Cloverleaf Amid AI Energy Crunch

Nvidia has acquired a minority stake in Cloverleaf, a 2024-founded startup that raised $300 million and acts as an intermediary between utility companies and data centers to arrange power and site infrastructure. The deal, reported on August 21, 2026, signals that grid capacity — not chip supply — has become the primary bottleneck constraining AI expansion. This follows a separate Nvidia-linked investment the same week into SB Energy's Ohio data center project, valued at $1.5 billion. By investing in the companies that power its GPU customers, Nvidia is effectively financing data center capacity ahead of demand, a model with precedent in the telecom industry. Analysts note this creates a feedback loop where Nvidia underwrites infrastructure that in turn generates GPU orders, with implications for the stability of current cloud compute pricing.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

BlocSignal Guide: Choosing Between Bloc, Cubit, and Signal in Flutter

A technical guide published on DEV Community outlines a decision framework for Flutter developers struggling to choose the right state management tool. The article introduces BlocSignal, a library that combines BLoC architecture with Preact Signals v7 to deliver synchronous, zero-millisecond state updates. It proposes a four-tier hierarchy ranging from raw signals for local widget state to HydratedMixin for persistence and undo-redo history. CubitSignal is recommended as the default for standard CRUD screens and form state, while BlocSignal is reserved for complex, event-driven pipelines requiring strict concurrency control. The guide aims to help developers avoid both over-engineering simple UI interactions and under-structuring complex transactional workflows.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DevOps Engineer Tejas Joins DEV Community to Share Cloud and K8s Insights

Tejas, a DevOps Engineer based in Bangalore, India, has introduced himself to the DEV Community platform as a new member. He specializes in AWS, Kubernetes, Terraform, and Jenkins in his day-to-day work. Tejas plans to publish content covering CI/CD pipelines, Kubernetes practices, and Infrastructure as Code patterns. He also intends to document real-world troubleshooting experiences to help others learn from common failures. He has invited fellow community members to connect and engage with his upcoming posts.

0 comments Read more at DEV Community