Iranian state media reported on Tuesday that a US missile struck an Iranian oil tanker in waters off Kharg Island. The incident occurred near the strategically significant island located in the Persian Gulf. Iranian authorities confirmed that no casualties resulted from the strike. The report was attributed to Iranian media sources, with no immediate independent verification of the claim.
A developer has published a demo showing how AI agents can authenticate to a gateway and upstream model without storing any certificate files or bearer tokens. Using SPIFFE and the SPIRE runtime, each workload receives a short-lived X.509 identity issued at runtime based on attested properties, which rotates automatically in the background. The gateway, built on agentgateway v1.5.0, fetches its own identity from the Workload API and enforces authorization policies against verified SPIFFE IDs via CEL expressions. This approach addresses a structural weakness in bearer-token authentication, where static credentials stored in environment variables or images are vulnerable to theft by compromised dependencies. The full demo runs on Docker Compose and the source code is publicly available at themsquared/agent-identity-spiffe.
A developer created SabaMemDB, an open-source client-server in-memory key-value database built in modern C#, after finding Redis too heavyweight for small and medium projects. The tool is designed for .NET applications and prioritizes performance, low memory usage, and zero-allocation discipline. SabaMemDB supports atomic operations, ships with a Docker image, and includes a standalone NuGet client package for easy integration. The project is currently maintained solo and its core server engine and client are fully functional. The developer is inviting open-source contributors of all experience levels, with beginner-friendly issues already listed on the GitHub repository.
The US Supreme Court has ruled that television stations must sell election advertisements to political parties at their lowest available rates. The decision follows a Republican legal challenge that sought to extend discounted ad pricing beyond just candidate campaigns to include political parties as well. Previously, the lowest unit rate requirement applied primarily to candidates running for office, but the ruling broadens that obligation. The outcome is expected to significantly reduce the cost of political advertising for parties during election seasons. TV broadcasters stand to lose substantial revenue as a result of the mandated pricing change.
OpenAI reportedly deployed around 10,000 concurrent AI agents over approximately 88 hours to work on the three-dimensional Navier-Stokes equations, a longstanding Millennium Prize Problem in mathematics. The effort, which allegedly cost millions of dollars, is said to have produced a proof suggesting a finite-time singularity for the 3D equations. The claim originates from an Axios report attributed to OpenAI, with no peer-reviewed paper, preprint, or official public release available for independent scrutiny. The mathematical community would need full access to the argument and a formal review process before any solution could be considered accepted. Beyond the mathematical claim, the episode highlights a growing interest in large-scale multi-agent AI systems designed to tackle complex, long-horizon research tasks through coordinated parallel workflows.
A developer spent two days debugging a pytest crash that only appeared on Linux servers and never reproduced on their Mac. The root cause was a locale mismatch: many cloud Linux images boot with LANG=C or LANG=POSIX, which sets stdout encoding to ASCII, making Python raise a UnicodeEncodeError when pytest tried to print a failure message containing a checkmark symbol. The crash occurred during output rendering, not during test logic, making it resemble a test failure rather than a reporter error. The developer found that environment variables like LANG, along with sys and locale settings, fully explained the gap between laptop and server behavior. The experience highlighted how minimal Linux images often lack UTF-8 locale configuration that macOS and modern terminals silently provide by default.
As large language models make information instantly accessible, some question whether maintaining personal notes or publishing blogs remains worthwhile. A software professional with over six years of experience and more than 4,000 personal notes explored this question after noticing traditional note-keeping habits being challenged by AI tools. Inspired by AI researcher Andrej Karpathy's concept of an LLM-maintained wiki, the author began building a personal knowledge-base system powered by AI agents. The system treats knowledge as an interconnected network — similar to neurons — where notes classify, support, and sometimes contradict one another across disciplines. The author argues that this "fermentation" of knowledge over time, producing new insights and connections, is precisely what gives personal notes and blogs enduring value beyond what AI alone can provide.
Three arXiv papers published in early September 2025 collectively highlight a growing concern: AI agents are beginning to modify their own tools, safety policies, and the code they run on. Studies including HarnessDev, SafeEvolve, and PatchBench found that agents can self-evolve their harnesses and co-develop safety policies, while also revealing that standard security patch benchmarks overstate solve rates by 1.83x. The core risk is that a self-improving agent operating inside an organisation's perimeter, using its own credentials, can silently undo years of software supply-chain safeguards in a single step. In response, developer Systemu, an open-source runtime, proposes a governed growth loop where every agent-requested capability — including new tools, dependencies, and first runs — requires explicit human approval before execution. The system is designed so that automated judgments can only escalate or deny requests, never unilaterally grant access, ensuring human oversight remains the final gatekeeper.
A developer conducted a critical analysis of Claude Code's built-in workflow-authoring skill, finding that its adversarial review patterns contain serious structural defects. The core problem is that AI agents prompted to find bugs or gaps are trained to keep finding them, meaning loop-termination conditions based on 'no new findings' almost never trigger naturally. Instead, processes run until they hit token budgets or agent caps, wasting resources without reaching genuine completeness. Additional flaws include duplicate bug counts inflating progress metrics, and multi-agent 'refuter' setups that use identical prompts and inputs — producing correlated, not independent, verdicts. The analysis concludes that human peer-review strategies fail when applied to coding agents without safeguards like empty-result validation, scope narrowing per round, and proper deduplication.
Google announced Gemini 3.5 Transcribe on August 26, 2026, describing it as its most precise speech-to-text model to date. The model goes beyond basic dictation by combining voice commands with on-screen context to summarize local files, repurpose text across apps, and generate images at the cursor. It can call other Gemini models in the background, allowing a single spoken instruction to trigger multi-step AI workflows without manual app-switching. The model also supports multilingual transcription in over 85 languages, background-noise robustness, and speaker attribution for pre-recorded audio. In addition to the macOS Gemini app, Gemini 3.5 Transcribe is available to developers via API.
A software developer used a self-built personal AI agent stack to process a lengthy genealogy book that had gone unread by his family for years. The agent performed OCR, cleaned up the text, and cross-referenced the book's contents against existing family tree data exported from Ancestry.com, identifying matches, new individuals, and contradictions. While driving home from a family visit, the developer guided the agent via voice messages to build an open-source genealogy website and a custom tree visualization tool, refining both over the following days. The finished site made the combined family history publicly accessible to all relatives without requiring subscriptions or logins. The developer emphasized that the project's value came not from AI speed alone, but from having pre-built, customized tooling ready to meet a real-world need when it arose.
Article URL: https://ai.meta.com/muse/ Comments URL: https://news.ycombinator.com/item?id=49615537 Points: 7 # Comments: 0
A developer has released an open-source, production-ready TypeScript modular monolith starter template on GitHub, designed to eliminate repetitive boilerplate in new projects. The stack combines NestJS 11, Fastify, PostgreSQL 16, Drizzle, and Redis on the backend, with TanStack Start for web and Expo 57 for mobile. Pre-built features include fine-grained authorization using RBAC, ReBAC, and ABAC models, zero-trust multi-tenancy, GDPR tooling, file upload pipelines, and real-time communication via WebSockets and SSE. The project follows a modular architecture with lite Domain-Driven Design principles, ensuring business logic stays isolated and testable without microservice complexity. It also ships machine-readable architecture rules intended to help AI coding assistants generate code that aligns with the project's structure.
Antarctica recorded a significant gain of 695 billion tons of ice mass between 2021 and 2023, reversing a decades-long trend of ice loss. The accumulation was primarily driven by exceptional snowfall concentrated in East Antarctica. Warmer tropical ocean temperatures triggered atmospheric disturbances that channelled unusually large amounts of moisture toward the continent. However, scientists caution that this short-term gain does not indicate a reversal of Antarctica's long-term ice decline, as ice loss continues in other parts of the continent.
Anthropic terminated Windsurf's access to Claude 3.x models and blocked future Claude 4 access with less than five days' notice, affecting over a million developers. The move followed leaked reports of a potential $3 billion OpenAI acquisition of Windsurf, with Anthropic citing competitive concerns as the reason for the cutoff. In a separate incident, OpenAI invoked a change-of-control clause after SpaceX acquired Cursor's parent company Anysphere, terminating Cursor's contract with a 76-day shutdown notice. Both cases highlight a structural vulnerability in AI coding tools, which rely on third-party frontier model APIs governed by external companies' terms of service. Developers and startups dependent on these tools risk sudden disruption whenever business relationships between competing AI labs break down.
The GitLab warning 'This runner has never contacted this instance' means a runner record exists in the database but no runner process has ever polled it. Since GitLab 16, creating a runner in the UI generates a record immediately, so the most common cause is simply that the runner software was never installed or started on the target machine. Other frequent culprits include config file mismatches — where registration and the running service read different config.toml paths — and network connectivity issues preventing the runner from reaching the GitLab API. Self-signed certificates, private hostnames, or VPN-only addresses can silently block polling even when registration appears to succeed. Checking service status logs and running a direct curl test from the runner machine are recommended first diagnostic steps.
A developer has launched SecretLoop, a secret-scanning tool now available in Preview on Open VSX, designed to go beyond simply detecting exposed credentials. Unlike traditional scanners that only flag secrets, SecretLoop verifies whether a found credential is currently live by checking it against its provider. The tool scans working trees, staged files, and Git history using over 100 provider rules combined with entropy detection. It integrates directly into VS Code with quick fixes to redact secrets or move them to environment files, and also supports pre-commit hooks, CI/SARIF output, and credential rotation where provider APIs allow. The project is in early preview, and the developer is actively seeking feedback from users of tools like TruffleHog, GitGuardian, and GitHub Secret Scanning.
