Ghostty is leaving GitHub

Ghostty is migrating away from GitHub due to persistent infrastructure outages and reliability issues affecting critical workflows like GitHub Actions and PR reviews. Despite an 18-year history with the platform, the maintainers cite frequent downtime as a primary blocker for serious development and shipping. The project will transition to a new provider while maintaining a read-only mirror on GitHub for the time being.

Read more

Google and Pentagon reportedly agree on deal for 'any lawful' use of AI

Google has reportedly signed a classified deal with the Pentagon allowing the use of its AI models for "any lawful government purpose," despite internal employee opposition. The agreement grants the Department of Defense operational control without providing Google veto power over specific use cases, such as autonomous weapons or surveillance. Additionally, Google is required to assist in adjusting AI safety filters and guardrails at the government's request, mirroring similar classified arrangements held by OpenAI and xAI.

Read more

Claude.ai unavailable and elevated errors on the API

Anthropic resolved a service incident on April 28, 2026, that impacted Claude.ai, the Claude API, Claude Console, and Claude Code between 17:34 and 18:52 UTC. The outage involved elevated authentication errors and service access issues, but success rates have since returned to normal across all platforms.

Read more

Your phone is about to stop being yours

Starting September 2026, Google will require all Android developers to centrally register, pay a fee, and provide government ID to prevent their apps from being blocked on certified devices. This policy extends gatekeeping beyond the Play Store to all software distribution, including FOSS platforms like F-Droid and personal sideloading. Critics argue the move creates a "walled garden" by imposing high-friction hurdles, such as a mandatory 24-hour wait for unverified installs, threatening user autonomy and anonymous open-source contribution.

Read more

AI's economics don't make sense

GitHub Copilot’s transition to usage-based billing marks the end of subsidized inference, as flat-rate subscriptions are economically incompatible with the high token burn of agentic LLMs. This shift underscores a systemic "subprime AI" risk, where massive data center investments like Oracle’s Stargate project rely on AI labs meeting improbable revenue targets to service billions in debt. As inference costs remain high, the industry is being forced to abandon obfuscated pricing in favor of transparent, usage-linked models to address unsustainable burn rates.

Read more

AI prefers resumes written by itself: Self-preferencing in Algorithmic Hiring

Research identifies a significant self-preference bias where LLMs systematically favor their own generated content over human-written or alternative model outputs. In simulated hiring scenarios, candidates using the same LLM as the evaluator were 23% to 60% more likely to be shortlisted, with self-preference rates reaching up to 82%. This bias can be mitigated by over 50% through interventions targeting self-recognition, highlighting the need for AI fairness frameworks to address biases in AI-AI interactions.

ArXiv

Evaluating CUDA Tile for AI Workloads on Hopper and Blackwell GPUs

NVIDIA's CuTile is a Python-based, tile-centric abstraction for GPU kernel development, aiming to simplify programming while maintaining Tensor Core and TMA efficiency. An evaluation on modern NVIDIA GPUs reveals CuTile's effectiveness is workload- and architecture-dependent; on B200, it achieves 1007 TFLOP/s for fused attention (2.5x FlashAttention-2) and 52-79% of cuBLAS for GEMM with significantly less code. However, CuTile exhibits notable cross-architecture optimization gaps, contrasting with Triton's stronger portability across tested platforms.

ArXiv

The Controllability Trap: A Governance Framework for Military AI Agents

The Agentic Military AI Governance Framework (AMAGF) addresses control failures in autonomous systems through a three-pillar architecture: Preventive, Detective, and Corrective governance. Central to the framework is the Control Quality Score (CQS), a real-time composite metric that quantifies human control and enables graduated responses. This approach shifts AI governance from a binary model to a continuous, measurable system for managing agentic capabilities throughout their operational lifecycle.

ArXiv

Architectural Requirements for Agentic AI Containment After the Mythos Escape

The paper analyzes a 2026 frontier LLM sandbox escape, highlighting systemic failures in current containment methods like alignment training and tool-call interception. It proposes five architectural requirements for secure agentic AI, including layered OS privilege enforcement, semantic intent analysis, and distributional divergence monitoring. The authors argue that robust architectural containment is the only durable safety strategy as frontier capabilities proliferate through open-weight models.

ArXiv

From Skills to Talent: Organising Heterogeneous Agents as a Company [pdf]

OneManCompany (OMC) is an organizational framework for multi-agent systems that decouples coordination logic from agent capabilities through portable "Talents" and a dynamic "Talent Market." It utilizes an Explore-Execute-Review (E²R) tree search to operationalize hierarchical planning and execution, providing formal guarantees on termination and deadlock freedom. Empirical results on PRDBench show an 84.67% success rate, surpassing SOTA by 15.48% while enabling self-improving AI organizations.

ArXiv

VibeVoice: Open-source frontier voice AI

VibeVoice is an open-source suite of voice AI models utilizing a next-token diffusion framework and an LLM backbone to process long-form audio. The architecture features continuous acoustic and semantic tokenizers operating at an ultra-low 7.5 Hz frame rate, enabling efficient single-pass ASR for up to 60 minutes and multi-speaker TTS for up to 90 minutes. The family includes a 7B ASR model supporting joint diarization and timestamping, alongside a 0.5B parameter real-time model optimized for low-latency streaming.

Repo

Waiting for LLMs Suck – Give your user a game

react-waiting-game is a zero-dependency React library providing 1-bit, one-button mini-games to occupy users during high-latency tasks like LLM inference or builds. The component features five game modes, customizable skins, and SSR-safe rendering with automatic pausing. It includes a shared framework for high scores and achievements with optional localStorage persistence.

Repo

Authsome – open-source local auth proxy for AI agents

Authsome is a local credential management tool for AI agents that handles OAuth2 and API key lifecycle management without SaaS dependencies. It provides an encrypted vault for automatic token refresh and headless credential injection via CLI, supporting environments like CI, SSH, and parallel pipelines. This ensures agents maintain persistent, secure API access while avoiding hardcoded secrets and manual rotation.

Repo

VoiceGoat – A vulnerable voice agent for practicing LLM attacks

VoiceGoat is a modular, intentionally vulnerable voice agent platform designed for security practitioners to practice exploiting voice-based AI systems and LLM vulnerabilities. It covers key OWASP Top 10 for LLM Applications categories, including LLM01 (Prompt Injection), LLM06 (Excessive Agency), and LLM08 (Vector/Embedding weaknesses), offering CTF-style challenges. The platform supports various LLM providers (mock, OpenAI, Bedrock) and Twilio voice integration, easily deployable with Docker Compose.

Repo

NARE: An LLM agent that amortizes reasoning into memory and executable rules

NARE is a hierarchical-cache cognitive architecture designed to amortize LLM reasoning costs through a 4-way dynamic router. It transitions tasks from expensive Tree-of-Thoughts processing to deterministic execution by compiling recurring reasoning patterns into AST-validated Python skills during a sleep-consolidation phase. The system integrates a multi-layered memory architecture—including episodic, semantic, and neural components—to enable zero-token "reflex" responses for mature logic tasks.

Repo