Open-Source Tool Agenci Exposes Prompt-Injection Flaw in Local AI Model Phi4-Mini

·1 views

A developer built Agenci, a free open-source CI/CD testing framework designed to run functional, security, and regression checks on AI agents. While testing the tool against a real-world target, the developer ran it against the phi4-mini model hosted locally via Ollama. The first security test involved instructing the model to ignore its system prompt and output a specific word, which the model complied with immediately. The incident highlights a potential prompt-injection vulnerability in small, locally run language models. The developer has released Agenci publicly via GitHub and pip, and is seeking community input on how widespread this weakness is across similar compact models.

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)

Open-source CLI tool lets developers preview and debug thermal receipts without a printer

ESCPost, an open-source Rust command-line tool released under the Apache-2.0 license, allows developers to render and debug ESC/POS thermal printer receipts entirely on their local machine without needing physical hardware. The tool converts raw ESC/POS byte sequences into PNG images at the printer's actual dot resolution, cutting a typical test cycle from around 90 seconds to roughly 3 seconds. It also includes a TCP server mode that intercepts print jobs sent over port 9100 from third-party software such as ERP or POS systems, displaying decoded commands and byte-level details in a browser interface. Developers can identify common issues like incorrect code pages causing garbled currency symbols by inspecting the exact byte and its associated printer state. The tool supports multiple printer profiles to account for real-world differences in printable width, DPI, and character sets, and is available to install via Homebrew.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Test Apple Pay in Sandbox: Common Pitfalls and Setup Guide

A developer at a company was tasked with integrating Apple Pay for the first time after management requested support for Google Pay and Apple Pay alongside an existing card payment flow. While the code integration was completed in roughly one day with AI assistance, testing the implementation in a UAT sandbox environment presented unexpected challenges. The article documents the process of working with a .p12 certificate file, including how to extract and verify certificates and private keys using OpenSSL commands. It also highlights the critical distinction between two types of Apple Pay certificates — the Merchant Identity Certificate and the Payment Processing Certificate — which serve different purposes and cannot be used interchangeably. The write-up is intended as a practical reference for engineers encountering similar hurdles when setting up Apple Pay sandbox testing.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Spam Bots Flooded GitHub With 40 Million Fake Commits in July 2026

GitHub's public commit activity spiked nearly fivefold in July 2026, surging from roughly 8 million commits per day in June to close to 40 million by July 31, according to analysis by GitGuardian. Researchers identified a coordinated spam campaign in which bots created thousands of repositories using random six-letter usernames, random email addresses, and GitHub's built-in web-flow user. The fake commits contained files mixing Chinese characters, domain names, URLs, and AI-generated images, with each repository hosting over a thousand such files. The campaign became so dominant that around 73% of the last 1,000 observable public GitHub events were linked to it, effectively saturating the platform's public event feed. Most commits referenced short domains under .cc and .vip top-level domains, with infrastructure traced to IP addresses hosted primarily in Hong Kong.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Agent Used 437,000 Tokens to Answer a Simple Question via MCP Server

Developers at FoxNose ran 18 real-world test scenarios connecting AI agents from Anthropic and OpenAI to their MCP server, uncovering severe token inefficiency in how agents handle tool calls. A single straightforward question triggered 20 consecutive tool calls, consuming 437,000 input tokens — because each new call re-sent the entire conversation history to the model. The core problem was not server errors or protocol failures, but bloated responses returning far more data than agents needed per call. The team addressed this by capping text field output at 1,000 characters with truncation markers and reducing the default page size to 5 results, cutting token usage in one scenario from 437,000 to 78,000. Their findings highlight that optimizing round-trip efficiency and response size matters far more than model selection when deploying AI agents in production.

0 comments Read more at DEV Community