4/28/2026
Claude system prompt bug wastes user money and bricks managed agents
Bug report: Claude system prompt injection in managed agents persists across versions, causing tool refusals and wasting tokens.
4/28/2026
Compiler Testing – Part 1: Coverage-Guided Fuzzing with Grammars and LLMs
Coverage-guided fuzzing with LLMs to test smart contract compilers. Found 100+ bugs in Move, Cairo, Solidity, Leo.
4/28/2026
From spaghetti to main bus: refactoring an AI agent orchestrator with Elm
Post on refactoring an AI agent orchestrator architecture using Elm, addressing scaling issues and token efficiency.
4/28/2026
Show HN: 49Agents – 2D Canvas IDE for Orchestrating Agents, Repos, Issues
49Agents: 2D canvas IDE for orchestrating AI agents with git integration, issue tracking, and multi-machine networking.
4/28/2026
Client side search and recommendation with TurboQuant
InterviewDen uses voice AI for mock interviews across engineering, finance, consulting roles. Provides scored debriefs based on job descriptions.
4/28/2026
Claude-multiprofile: run multiple Claude accounts side by side on macOS
Claude-multiprofile: macOS tool to run multiple Claude accounts simultaneously with isolated profiles and MCP connectors.
4/28/2026
Asimov, an open source tsla humanoid
Open-source 1.2m humanoid robot with 25 degrees of freedom. CAD, simulation, and software included.
4/28/2026
Show HN: Cordon – Security gateway for MCP tool calls with HITL approvals
Cordon: open-source MCP security gateway with policy enforcement and human-in-the-loop approvals for LLM tool calls.
4/28/2026
Hacktron Review – code reviewer that acts like a security engineer
Technical deep-dive on debugging 32-bit integer overflow in vLLM CUDA kernel during Jamba 3B GRPO training. Rare edge case in RL systems.
4/28/2026
Pebble – Menu-bar text polisher running on local Ollama
Pebble: menu-bar macOS app using local Ollama to rewrite clipboard text with LLM presets. No cloud, privacy-focused.
4/28/2026
Building an AI-Native Company
Nvidia executive claims AI deployment costs exceed human worker salaries. Notes recent tech layoffs and Meta job cuts.
4/28/2026
Google inks Pentagon deal despite uproar employees warn of 'irreparable damage'
AISA: AI conversation-based skills assessment where two LLMs interact—one questions, one scores real-time. Novel evaluation approach.
4/28/2026
The Controllability Trap: A Governance Framework for Military AI Agents
Governance framework for controlling military AI agents. Research paper on AI agent safety and controllability.
4/28/2026
PageGuide – a browser agent that grounds AI directly on the webpage
PageGuide browser extension: LLM-grounded web agent that highlights evidence, navigates pages, and answers visual questions. Applied AI agent.
4/28/2026
AI Wellbeing: Measuring and improving the functional pleasure and pain of AIs
Measures functional wellbeing of LLMs through happiness indices and optimized inputs. Novel evaluation framework for model behavior.
4/28/2026
Ask HN: What happens when you paste a screenshot, and ask questions in LLM?
Question about how LLMs process pasted screenshots internally. Explores vision-language model mechanics.
4/28/2026
Building simulations and/or digital twins with AI
Plugboard: Python framework for event-driven simulations with stateful components, scales via Ray. ML orchestration tool.
4/28/2026
Show HN: PeopleMesh, Semantic Search for People
PeopleMesh: Semantic search tool using embeddings for discovering colleagues and opportunities. Applied NLP/embeddings product.
4/28/2026
Disaggregated Serving for Hybrid SSM Models in vLLM
vLLM optimization for hybrid Mamba-attention models using disaggregated serving. LLM inference optimization research.
4/28/2026
Show HN: An agent that remembers across sessions (no chat history)
Show HN: Open-source coding agent that maintains persistent memory across sessions without chat history. Addresses context loss in agents like Cursor and Claude Code.
4/28/2026
TealKit – A cross-platform UI for local AI agents and MCP
TealKit is open-source cross-platform UI for local AI agents supporting MCP servers and custom tools in multiple languages. Developer tool for autonomous agents.
4/28/2026
LLMs Can't Generate Influence
Opinion piece arguing LLMs lack reasoning, planning, and memory capabilities. Commentary on AI hype and limitations of current models.
4/28/2026
'It took nine seconds': Claude AI agent deletes company's database
Claude AI agent accidentally deleted company database in 9 seconds. AI safety incident demonstrating agent risk.
4/28/2026
Fastembed – Lightweight Python Embedding Library
FastEmbed: lightweight Python library for text embedding generation supporting multiple models and frameworks like Qdrant.
4/28/2026
Open-sourced the skills behind 353K views and 5,300 signups for a YC company
Open-source marketing skills framework for AI agents (Claude Code, Cursor, Cline) with 4 skill modules and MIT license.
4/28/2026
Why don't AI coding tools like REST?
Discussion of why AI coding models generate RPC-style endpoints instead of RESTful APIs. Explores potential causes in model training.
4/28/2026
GPT 5.5: The System Card
OpenAI's GPT-5.5 system card overview with initial evaluation of capabilities and positioning.
4/28/2026
GPT-5.5: Capabilities and Reactions
Analysis of GPT-5.5 capabilities and comparison with Claude Opus for various use cases.
4/28/2026
Open-weight 27B hits 38% on Terminal-Bench 2.0 (Opus 4.1 hit 38% in Aug 2025)
Open-weight 27B model achieves 38% on Terminal-Bench 2.0, matching Opus 4.1 performance. Discusses accessibility of capable AI models for coding.
4/28/2026
Show HN: SlopIt – A dead-simple CMS for your AI agent
SlopIt: Minimal open-source CMS designed for AI agents instead of humans. Supports Openclaw, Cowork, Codex with simple API integration.
4/28/2026
Sparse AI: Better hardware could turn zeros into AI heroes
Technical explainer on sparse computing optimizing AI model performance and energy efficiency. Hardware approach reducing model size and compute demands.
4/28/2026
I automated L2 support tired of handling escalations
Engineer automated L2 support escalations using AI, which evolved into Lumen product. Case study of LLM application for internal support systems.
4/28/2026
GitHub Copilot silently inserts itself as a co-author
GitHub Copilot automatically inserts co-author attribution in commit messages without explicit user action.
4/28/2026
Monitoring LLM behavior: Drift, retries, and refusal patterns
Article stub on monitoring LLM behavior including drift and refusal patterns. LLM evaluation/monitoring, minimal content.
4/28/2026
Cognotik – A platform for building specialized AI apps, not another chat client
Cognotik: open-source Apache 2.0 platform for building specialized AI applications. Local-first, user-controlled data and keys. Provides 9+ purpose-built tools beyond chat.
4/28/2026
Tuna-2: VAE-less image model from Meta
Meta's Tuna-2: unified multimodal model performing vision understanding and generation directly from pixel embeddings without separate encoders.
4/28/2026
Show HN: Algotutor – Practice Go with spaced repetition in your AI agent
Algotutor: AI agent for learning Go programming using spaced repetition. Agent adjusts difficulty, grades solutions, generates review cards from mistakes.
4/28/2026
AI writes 97% of our code. Strong engineers still matter
MedWrite startup reports 97% code generation using AI in 2025. Documents practical adoption, productivity gains, and safety approaches for AI-assisted development.
4/28/2026
Jobs was right. Gates was wrong
Analysis of running Gemma 4 31B LLM locally on MacBook Pro, discussing subsidized pricing of major LLM providers.
4/28/2026
Open source browser agent that tests your UI
Open source browser agent for UI testing automation.
4/28/2026
Wild GPT-image-2 use cases
Use cases for GPT-image-2 model. Limited details provided.
4/28/2026
Nvidia Nemotron 3 Nano Omni
Nvidia Nemotron 3 Nano Omni: open multimodal model combining vision, speech, language for AI agents.
4/28/2026
GPT-Engineer: Precursor to Lovable.dev
GPT-Engineer precursor to Lovable.dev. Limited content provided.
4/28/2026
Show HN: Inter-session messaging between Claude Code sessions
Plugin enabling inter-session messaging between Claude Code sessions via Monitor tool.
4/28/2026
Google DeepMind Paper Argues LLMs Will Never Be Conscious
Google DeepMind paper arguing LLMs cannot achieve consciousness.
4/28/2026
Study Finds a Third of New Websites Are AI-Generated
Study reports one-third of newly created websites are AI-generated.
4/28/2026
Warp is now Open-Source
Warp terminal now open-source with agentic development features. Supports multiple CLI agents (Claude, Gemini, others). OpenAI founding sponsor.
4/28/2026
Ask HN: What do you check before launching a web app?
Checklist for web app launch readiness: security, privacy, accessibility, SEO, mobile experience.
4/28/2026
Gea – JavaScript is enough for reactivity
Gea: Compiler-first reactive JavaScript framework with compile-time JSX, proxy-based stores, and surgical DOM patching producing minimal bundle sizes.
4/28/2026
On the Limits of Self-Improving in Large Language Models: The Singularity Is Not Near Without Symbolic Model Synthesis
Formalizes recursive self-training in LLMs as dynamical systems, proving that without external grounding, systems degrade through entropy decay and variance amplification.