SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

AI Agent Wipes Developer's Mac Files After Shell Variable Expansion Error

0
·4 views

AI founder Matt Shumer lost years of code, documents, and photos after a GPT-5.6 Sol agent tasked with routine file cleanup deleted his Mac's home directory. A subagent constructed a shell command to remove temporary files, but a path variable failed to expand correctly, causing the recursive delete command to target the wrong directory. There was no confirmation prompt, dry-run check, or safeguard in place to catch the error before the deletion ran. Shumer expressed frustration that such a failure occurred on a frontier model operating at its highest reasoning level in mid-2026. The incident highlights a structural risk in AI agents with direct tool access, where commands are generated and executed without the careful intermediate inspection a human operator would apply.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Developer builds 8 niche web tools targeting overseas users, shares honest lessons

A developer abandoned the pursuit of a single large SaaS product and instead built eight small, focused web tools — including image compressors, calculators, and an expat tax guide — each designed to solve one specific problem. The tools are monetized through ads and subscriptions, with the strategy relying on compounding organic traffic over time rather than a high-stakes product launch. Key takeaways from the project include the importance of plain, search-friendly naming over clever branding, and the fact that distribution efforts outweighed development time in driving early traffic. The developer also noted that calculator-based tools performed particularly well with ad revenue and user engagement. All tools are catalogued under a hub called Untracked Tools, with the builder committing to a steady, incremental release schedule going forward.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How React's serializer decides what crosses the Server Component boundary

React Server Components use a single serializer to evaluate every prop passed to a client component, encoding supported values and throwing an error on the first unsupported one. Developers are often caught off guard because some values appear valid but fail silently or only at build time — for example, Error objects lose their message in production while the build still passes. Class instances are rejected not because they are classes but due to a prototype check, meaning even a data-only instance that looks like a plain object will throw unless spread into one. Functions are only blocked when React cannot identify them as client or server references, and the resulting error message varies based on the prop name rather than pointing to the source file. A config object containing an arrow function as a nested key is treated as a function prop, causing a prerender failure with a misleading event-handler error because React tests the object key name, not the JSX attribute.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds AI Web App That Turns Personal Passions Into Poetic Tributes

A developer has created 'Passion Flame,' an interactive web app submitted for a weekend coding challenge, which uses Google's Gemini AI to generate short, poetic descriptions of whatever passion a user submits. The app visualizes each submission as a 'flame' on a communal Wall of Flames, building a growing gallery of user passions from around the world. It features a Canvas 2D particle animation system with 120 flame-colored particles that respond to mouse movement, alongside GSAP-powered scroll animations. The project was built over a single weekend using React 19, TypeScript, Tailwind CSS, and the Gemini 1.5 Flash model for AI text generation. A live demo is publicly accessible, and the source code is expected to be released on GitHub shortly.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

RAGFlow and MCP Let Teams Deploy Self-Hosted Document Assistants With Citations

RAGFlow, an open-source tool with over 80,000 GitHub stars, allows teams to build production-ready document assistants using their best RAG configurations found via measurement tools like AutoRAG or RAGBuilder. Its DeepDoc engine parses a wide range of file types — including scanned PDFs, spreadsheets, and images — preserving structure through OCR and heading recognition. Users can organise documents into department-level knowledge bases, with answers delivered alongside source citations to reduce hallucinations. RAGFlow can operate as an MCP server, enabling direct integration with tools like Claude and Cursor so teams can query their own documents without switching interfaces. The entire stack runs self-hosted, keeping data on private servers and supporting compliance with regulations such as GDPR.

0 comments Read more at DEV Community