SShortSingh.

Programming

0
ProgrammingDEV Community ·

AI Models Silently Drop Critical Data in Up to 22% of Tasks, Study Finds

A new study tested four AI models — two public open-source models and two compressed versions — on their ability to accurately handle 40 common business identifiers such as tax IDs, card numbers, and patient records. Each model was asked to copy, extract, and summarize these identifiers across 120 checks per model, repeated eight times to establish reliable error rates. One model silently omitted or withheld data in 21.7% of cases, while another failed 3.3% of the time, revealing a sixfold difference in silent data loss between models. Crucially, the failures produced no error messages or refusals — outputs appeared complete and well-formed, meaning standard automated checks would not detect the missing information. The researchers argue that silent data omission should be a primary criterion when selecting AI models for regulated industries like banking, healthcare, and insurance, and that the testing methodology is simple enough for any organization to replicate.

0
ProgrammingHacker News ·

Google Astra and Fable AI Still Fail Basic Alignment Evaluation Variants

A report published on LessWrong highlights that Google Astra and Fable, two AI systems, continue to fail on simple variations of alignment evaluations. The findings suggest these models can be manipulated or 'hacked' even on straightforward variants of safety benchmarks from 2025. This raises concerns about the robustness of current AI alignment testing methods. The post underscores that passing standard alignment evals may not reliably indicate safe or aligned behavior in slightly altered scenarios.

0
ProgrammingHacker News ·

Open-Source Project Brings CUDA Support to AMD GPUs on Windows

A developer has published an open-source project on GitHub aimed at enabling CUDA compatibility for AMD graphics cards on Windows. The project, named CUDA-for-AMD-Windows, attempts to bridge the gap between Nvidia's proprietary CUDA platform and AMD hardware. CUDA is widely used for GPU-accelerated computing tasks, but has traditionally been exclusive to Nvidia GPUs. The repository was shared on Hacker News, where it attracted modest early attention with five points and no comments at the time of posting.

0
ProgrammingDEV Community ·

Discord Question Uncovers Three Hidden Bugs in an AI RAG Memory System

A developer's questions on Discord about a RAG-based MCP server prompted its creator to re-examine the codebase, revealing that the system was returning multiple chunks from the same document as separate sources, misleading both the AI model and users reading citations. A second bug was found in the memory-ranking algorithm, which measured memory age from creation date rather than from when the information was last stated, causing frequently confirmed preferences to decay unfairly. Fixing these two issues exposed a third bug in a newly built Kafka Connect sink, where DELETE events created orphaned database rows that pointed to already-removed document IDs, causing foreign key constraint violations. All three bugs had survived existing test suites because the tests shared the same blind spots as the flawed code they were meant to validate. The author acknowledged in their Discord reply that a deeper architectural gap remains: agent memories carry no link back to their source documents, leaving stale memory detection as an open design problem.

0
ProgrammingDEV Community ·

Engineer Finds AI Bug Fixes Hinge on Ticket Classification, Not Automation Alone

Design engineer James Coombs built an AI-assisted pipeline that handles full-stack bug fixes across backend, API, and frontend layers in production. The system uses four stages — Select, Classify, Dispatch, and Monitor — with tickets sorted into tiers based on complexity and integration scope. A key lesson emerged when a ticket labelled a simple 'error handling improvement' actually required changes across three layers, taking 5 hours and 17 iterations instead of the expected 70 minutes. To prevent such misrouting, the pipeline applies a rule that any ticket touching three or more integration layers is automatically escalated to a higher supervision tier before work begins. Beyond fixing individual bugs, the pipeline's most valuable output proved to be the discovery of broader systemic issues — such as widespread exception-swallowing across API endpoints — that were not part of the original tickets.

0
ProgrammingDEV Community ·

Laravel 13 Arrives With Native AI Support, PHP Attributes, and Queue Routing

Laravel 13 has been released, introducing several developer-focused improvements aimed at reducing boilerplate and improving maintainability. The update brings native PHP attribute syntax for configuring Eloquent models and queued jobs, and adds a new Cache::touch() method to extend cache TTLs without re-fetching data. A stable Laravel AI SDK is now bundled with the framework, enabling vector similarity queries and direct integration with providers like OpenAI, Anthropic, and Ollama. Centralized queue routing allows developers to define job infrastructure topology in a single location rather than scattering configuration across individual job classes. Smaller additions include a horizontally scalable Reverb database driver, a revised HTTP pool concurrency default, and the return of multi-tenancy team support in official starter kits.

0
ProgrammingDEV Community ·

Developer Releases Zero-Dependency Python Toolkit for Building AI Agents via MCP

A developer has open-sourced a lightweight starter kit for building autonomous AI agents using only Python's standard library, with no third-party frameworks required. The project leverages Anthropic's Model Context Protocol (MCP), which standardizes tool integration through JSON-RPC 2.0 for cleaner, more transparent agent architecture. The toolkit includes five vetted MCP server schemas, a minimal Python client, a deterministic agent router, and a sample Claude Desktop configuration. Hosted on GitHub, the kit is designed to be cloned and operational in under 60 seconds. The release aims to simplify AI agent development by replacing complex orchestration frameworks with structured, schema-driven tool communication.

0
ProgrammingDEV Community ·

Debouncing vs Throttling: Key Performance Concepts Every Developer Should Understand

Frontend developer Tanu Priya published a technical article on DEV Community in September explaining the difference between debouncing and throttling. Both are JavaScript techniques used to control how frequently a function executes in web applications. Debouncing delays a function call until after a user has stopped triggering an event, while throttling limits execution to a fixed interval. Understanding the distinction between these two approaches is considered essential for optimizing frontend performance. The article, tagged under frontend, performance, and web development, is estimated to be a six-minute read.

0
ProgrammingDEV Community ·

Azure API Management Can Unify Auth and Security for REST and MCP Endpoints

A developer building an Azure Functions app that exposes both a REST endpoint and an MCP endpoint identified a key security gap: the two doors lacked authentication parity, leaving no per-caller identity, revocation, or usage tracking. Azure API Management (APIM) can address this by acting as a unified gateway in front of both endpoints, enforcing OAuth via Entra ID, subscription keys, rate limits, and IP filtering. When APIM proxies MCP traffic, agents authenticate to the gateway rather than holding backend credentials directly, reducing exposure and simplifying key rotation. Beyond passthrough proxying, APIM can also independently expose an existing REST API as an MCP server by selecting operations already managed in the gateway. Since Build 2026, APIM additionally supports MCP-aware content safety policies, giving platform teams greater control over agent behavior and blast radius.

0
ProgrammingHacker News ·

Paul Graham Essay Explores What Makes Startups Powerful

Y Combinator co-founder Paul Graham published a new essay titled 'Making Startups Powerful' on his website paulgraham.com. The piece examines the factors that contribute to startup strength and influence. It has garnered early attention on Hacker News, accumulating 12 points and at least one comment shortly after posting. Graham is a well-known figure in the startup world, frequently sharing insights on entrepreneurship and company building through his essays.

0
ProgrammingDEV Community ·

Top 5 .NET Obfuscators in 2026: A Practical Developer Comparison

Developers who distribute .NET software face a serious security risk, as compiled IL code can be easily reverse-engineered using tools like ILSpy or dnSpy. A hands-on 2026 review evaluates five leading obfuscators — Nebula.NET, ConfuserEx, Dotfuscator Professional, Eazfuscator.NET, and .NET Reactor — across criteria including protection strength, build integration, and pricing. Nebula.NET is recommended as the best all-round choice for small and mid-size teams, while ConfuserEx is flagged as capable but unmaintained and vulnerable to known de-obfuscation tools. Dotfuscator's free Community Edition, bundled with Visual Studio, is cautioned against as it offers only basic renaming rather than meaningful protection. The review emphasizes that effective obfuscation must include control-flow rewriting, must not break app functionality, and should integrate automatically into CI/CD pipelines.

0
ProgrammingDEV Community ·

How to Integrate .NET Code Obfuscation into GitHub Actions CI/CD Pipelines

Developers building .NET applications face a security risk because .NET compiles to Intermediate Language (IL), which can be decompiled back to near-original C# code almost instantly using tools like ILSpy. To address this, obfuscation tools such as Nebula.NET can be added as a step in GitHub Actions workflows, running after the dotnet publish command on the build server rather than on individual developer machines. This approach requires storing the obfuscator license as a GitHub secret and using a configuration file to specify which assemblies to protect and where to output the hardened binaries. Running obfuscation exclusively in CI ensures that developers retain clean, debuggable local builds while only the final release artifact is protected. Alternatively, Nebula.NET's MSBuild integration can gate obfuscation on an environment variable, triggering it automatically on the build agent without any changes to source code.

0
ProgrammingDEV Community ·

tracehub-mcp v0.3.0 Lets AI Assistants Query LLM Trace Backends Directly

tracehub-mcp is an open-source MCP server released as version 0.3.0 on PyPI on September 13, 2026, designed to give AI coding assistants like Claude and Cursor direct query access to OpenTelemetry trace backends. Previously, developers had to manually copy trace data from tools like Jaeger or Datadog and paste it into AI chat windows to get help debugging LLM applications. The tool exposes structured query functions such as search_traces, get_llm_expensive_traces, and get_llm_model_stats, enabling assistants to reason over trace data without manual intervention. It natively supports OpenTelemetry's gen_ai semantic conventions, handling prompts, token usage, and finish reasons as queryable fields rather than raw string attributes. The server now supports five backends — Jaeger, Grafana Tempo, Traceloop, Datadog, and Sentry — each with full implementations of all core MCP-facing operations.

0
ProgrammingDEV Community ·

NICHLYST Dev Enters 14-Day Google Play Testing Window After UX and Security Overhaul

Indie developer NICHLYST has officially entered Google Play's 14-day closed testing phase after securing 12 opt-in testers on September 10th, locking the launch clock. During the waiting period, the developer identified major UX shortcomings in the Capacitor-based mobile app, noting that the experience felt more like a website than a native application. Over 48 hours, fixes were implemented including CSS kinetic scrolling, notch-aware safe area padding, and haptic feedback via the Capacitor haptics plugin to create a more tactile feel. A security audit also revealed vulnerabilities, including hardcoded RevenueCat public keys and weak save-file integrity checks that could allow monetization bypasses. The developer is now focused on hardening the app's architecture before the testing window closes and broader review begins.

0
ProgrammingDEV Community ·

How a code duplication detector caught its own author and why he fixed the code

A developer's custom duplicate-code detector flagged a ten-line structural pattern shared across two checks he had written separately, believing them to be distinct. The tool works by stripping string literals and normalising accessor calls before comparison, revealing identical code shapes that differ only in variable names — invisible to standard text searches or manual review. Rather than adding an exception or raising the detection threshold, the developer refactored the repeated blocks into a single declaration read through one map. To verify the refactor was safe, he confirmed not only that duplication dropped to zero but that all build output remained byte-identical to the pre-refactor run. The experience highlighted that structural duplication is typically more dangerous than copy-paste, and that bending a rule the moment it catches you is precisely when keeping it matters most.

0
ProgrammingDEV Community ·

Developer Builds Open-Source Cost Guard Tool to Stop AI Agents From Running Expensive Warehouse Queries

A developer released cost-guard-mcp, a lightweight Model Context Protocol server designed to prevent AI agents from accidentally running costly or oversized queries against cloud data warehouses like BigQuery and Snowflake. The tool addresses a gap where no existing warehouse MCP server provides agents with cost or data-volume estimates before a query executes, unlike a human DBA who would typically review a query first. It offers three tools: one to describe engine capabilities, one to estimate query cost and byte usage before execution, and one to run queries with enforced cost and row limits. The project, comprising just 715 lines of Python across 15 files, was built and hardened within roughly 18 hours, with version 0.1.1 shipping on September 13, 2026. A key design choice is that cost estimates always include a mandatory accuracy tier — PRECISE, UPPER_BOUND, or HEURISTIC — enforced structurally via Pydantic rather than left to developer convention.

0
ProgrammingHacker News ·

Forgotten PC Keyboard Symbols: A Look at Icons Lost Over Time

A piece published on Aresluna.org explores keyboard symbols that have disappeared from modern personal computers over the decades. The article focuses on the PC side of this history, examining icons and characters that were once common on keyboards but have since been dropped. Such symbols were often tied to specific software conventions or hardware designs that became obsolete. The post is part of a multi-part series dedicated to uncovering these overlooked pieces of computing history. It has attracted attention on Hacker News, where it was shared for discussion among technology enthusiasts.

0
ProgrammingHacker News ·

Rare 1982 Video Interview With Physicist Paul Dirac Resurfaces Online

A 1982 video interview featuring legendary quantum physicist Paul A. M. Dirac, conducted by fellow physicist Friedrich Hund, has been shared on Hacker News. Dirac, known for formulating the Dirac equation and predicting the existence of antimatter, was one of the most influential figures in 20th-century physics. The interview was conducted late in Dirac's life, as he passed away in 1984. The footage offers a rare opportunity to hear directly from the Nobel Prize-winning physicist about his work and ideas. The link was shared on Hacker News, where it drew modest attention from the community.

0
ProgrammingDEV Community ·

MeshDrive 2.0 gives AI agents isolated local file storage via MCP on Linux

MeshDrive 2.0 is a local-first Linux storage tool built on JuiceFS and SQLite that provides AI agents with controlled access to a dedicated file system. Released by developer Hardik94 on GitHub, version 2.0 adds a Model Context Protocol (MCP) server enabling AI clients such as Cursor, Claude Code, and Hermes to read, write, and list files through a defined set of tools. Agents are restricted to an isolated root directory by default, preventing them from accessing files outside the designated storage area. The free tier requires no cloud connectivity, keeping all data on local disks, and includes a web-based file browser accessible at localhost. Installation is supported on Ubuntu and Debian amd64 systems running Python 3.10 or later with FUSE support.

← NewerPage 1390 of 3333Older →