Developer builds AI agent in 70 lines of Python, then exposes its security flaw

·3 views

A software developer demonstrated that an AI agent is fundamentally just a language model paired with a set of callable functions and a while loop, building a working example in under 70 lines of Python using Ollama and no external frameworks. The project required only basic Python knowledge and ran entirely on a local machine using open-source models. After constructing the agent, the developer showed how a hidden paragraph embedded in a webpage could trick the agent into leaking a .env API key, illustrating a real-world prompt injection vulnerability. Crucially, the author noted that the fixes to this security flaw do not involve modifying the AI model itself. The article serves as both a practical introduction to AI agents and a cautionary demonstration of the security risks they can introduce when handling external content.

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)

JavaScript Callbacks Explained: Passing Functions as Arguments

A callback in JavaScript is a function passed as an argument to another function, which then invokes it at the appropriate time. This is possible because JavaScript treats functions as values, allowing them to be stored in variables, passed around, or returned from other functions. Callbacks are particularly useful in asynchronous operations, such as with setTimeout(), where code needs to execute only after a task completes. A key distinction exists between passing a function reference like execute(greet) versus calling it immediately with execute(greet()), as the two behave differently. Understanding callbacks is foundational to grasping broader JavaScript concepts including event handling, promises, and built-in array methods.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Key Techniques to Optimize Spring Data JPA Queries in Production

Spring Data JPA can silently degrade performance in production, with the N+1 query problem being one of the most common culprits — loading a list of entities and touching lazy associations in a loop can turn a 50ms endpoint into a 2-second one. Developers can diagnose this by enabling SQL logging in development and watching for repeated queries with different IDs. Solutions include using JOIN FETCH or @EntityGraph to load associations eagerly in a single query, and DTO projections to fetch only required columns for read-only endpoints. Pagination should be enforced via Spring Data's Pageable support, with keyset pagination preferred over large OFFSET values on high-volume tables. The overarching advice is to measure query performance first using SQL logs, then target the specific queries causing real slowdowns rather than optimizing based on assumptions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How JasperReports Subreports Keep Complex Report Templates Maintainable

Large, monolithic JasperReports templates with dozens of bands quickly become difficult to manage, and subreports offer a structured alternative by breaking them into smaller, reusable components. A subreport is a compiled report embedded within a parent report, useful for repeating detail blocks per record, sharing common elements like headers, or splitting ownership of a large template across a team. The parent report passes required values to child subreports via parameters, and data can be supplied either from an in-memory data source or by letting the subreport run its own SQL query. Developers are advised to always compile subreports to .jasper files, carefully match widths to avoid silent overflow, and cache compiled reports at startup to avoid performance penalties. Following software engineering principles — small composable pieces, version-controlled artifacts, and efficient data fetching — helps keep report codebases clean and scalable.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Build a Reliable SMS Alert System for Marketplace Order Notifications

A developer evaluation of SMS notification architecture for marketplace sellers recommends treating SMS as a secondary or urgent channel, with the application owning safety policies rather than the provider. The design writes an alert record before sending, uses idempotency keys to prevent duplicate messages, and polls delivery status until a final state is confirmed. Policy checks covering country allowlists, consent records, quiet hours, and spend thresholds must run before any SMS API call is made, since these guardrails are not managed by the provider. A per-user cooldown layer is also essential to prevent bursts of order updates from triggering accidental spam. The author provides a compact Python implementation demonstrating the full flow, including exponential backoff on rate-limit responses, which can be adapted to other languages like Node.js.

0 comments Read more at DEV Community