Automated AI News Brief: OpenAI Solves PHP2Shell Zero-Days with GPT-5.6 Sol, Debuts Decisions API, NVIDIA Releases OpenShell, and Anthropic Donates Agent Skills

Introduction

Horizon gathered the material for this post, and Codex selected and rewrote it; Horizon is used only for data collection. Today's technical developments reflect six major trajectories across the AI landscape: AI-driven cyber defense pioneering the "Defender's Window," kernel-level sandbox isolation for autonomous agents, latency micro-optimizations in bounded decision APIs, agent procedural standards transitioning to vendor-neutral foundations, reinforcement learning guided by real-world system side effects, and extreme low-precision edge audio inference. From OpenAI and independent researchers using GPT-5.6 Sol and Gemini 4 Argon to preemptively identify catastrophic PHP core engine zero-days (PHP2Shell) powering 70% of the internet, to NVIDIA open-sourcing OpenShell for containment of multi-agent fleets, and Anthropic officially handing the Agent Skills specification to the Linux Foundation, AI systems are expanding from code generation into systems-level security, formally verified autonomous agents, and universal open standards.

PHP2Shell Zero-Day Disclosures: OpenAI's Patch the Planet and GPT-5.6 Sol Uncover Core PHP Vulnerabilities Affecting 70% of the Web

In software infrastructure security, a coordinated defense effort spearheaded by frontier AI reasoning models has been declassified. Security researchers have revealed two critical memory-corruption vulnerabilities buried deep within the core PHP engine, designated CVE-2026-109618 and CVE-2026-109619 and collectively dubbed "PHP2Shell." Because more than 70% of web servers—including major CMS ecosystems such as WordPress, Drupal, and Joomla—rely on PHP, participating researchers characterized the flaw as one of the most consequential vulnerabilities ever coordinated.

Crucially, the vulnerabilities do not originate in third-party plugins or web application authentication logic. Instead, they reside in PHP's internal memory management routines when handling out-of-memory (OOM) conditions and fatal errors. Crafted requests can manipulate these conditions to achieve unauthenticated remote code execution (RCE) alongside Heartbleed-style arbitrary memory disclosure across affected deployments.

This discovery represents a landmark demonstration of what OpenAI terms the "Defender's Window"—the capacity for advanced reasoning models to uncover deep structural vulnerabilities before malicious threat actors can weaponize them. The bugs were uncovered by researchers at Calif using OpenAI's GPT-5.6 Sol reasoning model, alongside independent discoveries and exploit-chain validation by Wiz utilizing Gemini 4 Argon. Over an intensive two-month confidential embargo, OpenAI coordinated with the PHP Foundation, national security agencies, and cloud infrastructure providers under its "Patch the Planet" initiative and Daybreak research access program. Cloudflare deployed universal WAF mitigations covering all tiers ahead of the public advisory, and OpenAI distributed a specialized Codex diagnostic skill to assist system administrators in auditing their deployments. Although an accidental community leak truncated the embargo hours before the official release, the incident demonstrates that frontier models can actively secure mission-critical open-source infrastructure at internet scale.

Source: [AINews] OpenAI’s Patch the Planet program is 1 of 2 parties to discover PHP2Shell vulnerabilities in core PHP affecting 70% of the web

OpenAI Introduces Decisions API: Powered by GPT-6 Luna for 10x Faster Bounded Decisions and Agent Routing

OpenAI has officially launched the "Decisions API," a specialized endpoint engineered to handle high-frequency, structured judgment tasks with sub-second latency and minimal token overhead. Powered by OpenAI's specialized compact model GPT-6 Luna, the Decisions API differs fundamentally from open-ended conversational endpoints by directing model intelligence strictly toward bounded questions with finite, predefined outcomes.

Key architectural highlights of the Decisions API include:

1. Ultra-Low Latency & High Throughput: Benefiting from targeted distillation and operator specialization, the endpoint operates up to 10x faster than traditional generation pipelines, making it suited for tight real-time interaction loops and latency-sensitive production services.
2. Three Core Judgment Primitives:
   • Predicate: Evaluates a contextual proposition and returns an exact probability score or boolean value determining truth.
   • Choice: Selects the optimal matching category from a mutually exclusive, predefined enumeration pool.
   • Score: Ranks or scores text, image inputs, or structured JSON against user-defined rubrics and continuous scales.
3. Massive Batch Capacity: Developers can evaluate up to 400 discrete decision questions in a single API round-trip, with native support for Prompt Caching, GPT-Live streaming hooks, and strict data residency compliance.

For software engineers, the Decisions API sets a new standard for automated customer support triage, real-time trust-and-safety screening, and autonomous agent next-hop routing where latency and token frugality are paramount.

Source: Build more responsive apps with the Decisions API

NVIDIA Open-Sources OpenShell: A Kernel-Level Secure Sandbox and Credential Firewall for Autonomous AI Agents

As autonomous coding agents and terminal-enabled tools become ubiquitous in engineering environments, organizations confront a fundamental dilemma: agents require access to local filesystems, package managers, command-line interfaces, and remote APIs to perform meaningful work, yet granting such permissions risks catastrophic data destruction and credential harvesting. To resolve this standoff, NVIDIA has released OpenShell—an open-source security runtime engineered specifically for fleets of autonomous AI agents.

OpenShell is built on the premise that agents should possess comprehensive execution capabilities without receiving unrestricted access to underlying infrastructure, secrets, or internal networks. Its defense architecture rests on three pillars:

1. Kernel-Level Policy Enforcement: Every agent executes inside an isolated micro-sandbox. OpenShell instruments the host kernel to monitor and enforce strict rules over every file access, system call (syscall), and outbound network connection. Unauthorized socket openings or attempts to touch protected directories are terminated directly at the kernel boundary.
2. Dynamic Credential Masking: Agents never handle authentic API keys, cloud access tokens, or SSH private keys. OpenShell functions as an intelligent egress gateway, dynamically injecting real credentials into outbound HTTP/gRPC headers only when requests align with verified policy rules directed to approved external endpoints.
3. Formally Verified Policy Evolution: When an agent requests elevated privileges during execution—such as communicating with a newly discovered domain or installing unexpected system dependencies—OpenShell's formal verification engine (Policy Prover) analyzes the proposed policy delta, ensuring high-risk capability expansions are quarantined for human engineer sign-off.

OpenShell supports Linux, macOS with Apple Silicon, and WSL 2, integrating with standard container runtimes including Docker and Podman to provide a production-ready security foundation for autonomous agents.

Source: OpenShell – A safe, private runtime for fleets of autonomous AI agents

Milestone for Agent Standards: Anthropic Donates Agent Skills to the Linux Foundation's AAIF, GitHub Adds Skills and MCP to Metrics API

The push toward open, vendor-neutral standards for autonomous AI agents has achieved a major milestone. Anthropic has officially donated the "Agent Skills" open standard—the directory structure centered on SKILL.md documents and executable procedural scripts that supply agents with domain-specific knowledge—to the Agentic AI Foundation (AAIF) under the Linux Foundation.

Originally conceived during the development of Claude Code to equip agents with reproducible engineering workflows, the Agent Skills format has rapidly become an industry standard. It is now supported by more than 70 agent platforms—including Cursor, GitHub Copilot, ChatGPT, and VS Code—with approximately 1 million public skills published and nearly 280 million installs recorded worldwide.

By donating the specification, trademark, GitHub organization, and the agentskills.io domain to the AAIF (which Anthropic co-founded with OpenAI and Block, and which already hosts the Model Context Protocol / MCP), Agent Skills transitions from a corporate project into an open digital public good maintained collaboratively by engineers from Google, OpenAI, Anthropic, and independent contributors. Upcoming governance milestones include semantic versioning, a formal RFC process, and structured security auditing guidelines.

In parallel, GitHub announced an update to its Copilot Usage Metrics API: enterprise administrators can now programmatically track individual consumption metrics for custom skills, bespoke agents, MCP servers, slash commands, and plugins created within their organizations, replacing generalized aggregates with granular telemetry.

Source: Donating Agent Skills to the Agentic AI Foundation / Custom skills, agents, and MCP servers in usage metrics API

Microsoft and Hugging Face Debut ThinkingBox-Training: Turning Database States into RL Rewards for Dependable Agents

In contemporary AI agent development, a pervasive evaluation blind spot exists: benchmarks frequently judge models based on the superficial syntactic plausibility of their dialogue or tool invocations. An agent can emit convincing explanations and report that a problem has been resolved, while the underlying enterprise database reflects an incomplete or erroneous state (such as marking a support ticket closed while shipping exception flags remain unhandled).

To address this failure mode, Microsoft and Hugging Face have introduced "ThinkingBox-Training." Building upon ThinkingBox-Bench—a benchmark spanning 507 multi-turn enterprise tasks across retail, supply chain, and administrative systems—this new framework transforms terminal database validation checks into reward functions for post-training reinforcement learning (RL).

The researchers discovered that conventional training sets frequently suffer from tasks that are either trivially solved or insurmountable for a given model. ThinkingBox-Training introduces an "Offline Policy-Relative Difficulty" algorithm to select tasks that fall precisely within the model's zone of proximal development. Applying Group Relative Policy Optimization (GRPO) over just 50 steps on Qwen3.8-27B lifted average pass@1 success from 51.70% to 60.78%. More importantly, the number of tasks completed flawlessly across all 20 consecutive runs surged nearly threefold, from 38 to 107. The project offers a practical blueprint for grounding agent post-training in tangible backend side effects.

Source: Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox

KittenML Releases Kitten-TTS 2: World's First 1.58-Bit Ternary TTS Model with Real-Time CPU Inference and Voice Cloning

In local and edge AI computing, ultra-low-precision quantization has achieved a breakthrough in speech synthesis. KittenML has open-sourced "Kitten-TTS 2" on GitHub and Hugging Face—the world's first state-of-the-art text-to-speech model built on a 1.58-bit ternary weight architecture (-1, 0, +1).

Featuring 1.7 billion parameters, Kitten-TTS 2 delivers several key technical capabilities:

1. Real-Time CPU-Only Inference: Because every linear layer utilizes ternary weights with a scale block size of 128, the model bypasses requirements for dedicated GPUs. Powered by optimized ternary CPU kernels, standard consumer processors achieve faster-than-real-time audio synthesis at minimal compute cost.
2. Zero-Shot Voice Cloning: Given fewer than five seconds of clean reference audio, the model captures speaker identity, vocal timbre, accent nuances, and acoustic acoustics without task-specific fine-tuning.
3. Multilingual Synthesis & Expressive Delivery: Supporting more than 20 languages with cross-language cloning, Kitten-TTS 2 provides granular emphasis and emotional control, enabling developers to steer prosody, stress specific syllables, and modulate expressive delivery programmatically.

Kitten-TTS 2 presents an open, self-hostable alternative to proprietary voice synthesis APIs for resource-constrained local agents and embedded hardware.

Source: Reddit LocalLLaMA: Kitten-TTS 2: New Open-source SOTA 1.58 bit TTS

DeepSeek and MiniMax Clear China's AI Export Regulations for Global Open-Source Distribution

In international AI regulatory policy, prominent Chinese foundation model developers DeepSeek and MiniMax have completed the administrative registration and cross-border security review required under China's interim AI export control regulations.

Administered by the Ministry of Commerce to oversee the international distribution of advanced technologies, the regulatory clearance provides explicit authorization for both companies to continue open-sourcing model weights internationally and pursuing authorized commercial cross-border partnerships. The clearance resolves regulatory ambiguity surrounding cross-border model weight distribution, ensuring that global developers and enterprise teams can continue incorporating future open-weights releases within a compliant legal framework.

Source: Reddit LocalLLaMA: DeepSeek and MiniMax Clear China’s AI Export Rules for Overseas Open-Source Releases

Daily Perspectives

Today's developments underscore a decisive shift in artificial intelligence from isolated generation to systems-level resilience:

1. AI Safety Transitions from Code Generation to Proactive Defense: The discovery of PHP2Shell highlights a tangible reversal in cybersecurity dynamics. Historically, systemic zero-days in ubiquitous open-source code were frequently hoarded by attackers. Reasoning models such as GPT-5.6 Sol and Gemini 4 Argon identifying memory-corruption flaws two months ahead of exploitation showcases the defensive leverage modern frontier AI provides to open-source infrastructure.
2. Agent Architecture Demands Kernel-Level Isolation and Open Standards: As autonomous agents assume greater operational agency, conversational safety guardrails prove insufficient. NVIDIA's OpenShell addresses this through OS kernel monitoring, dynamic credential proxying, and formal verification. Simultaneously, Anthropic's donation of Agent Skills to the Linux Foundation's AAIF ensures the protocols governing agent procedures remain open, neutral, and cross-platform.
3. Defining Agency Through Concrete State and Accessible Inference: Microsoft's ThinkingBox-Training reinforces an essential insight: agent competence must be validated by the actual state of underlying databases rather than conversational output alone. Combined with architectural breakthroughs like Kitten-TTS 2—which delivers expressive voice cloning on commodity CPUs via 1.58-bit weights—the trajectory of AI points toward verifiable reliability and lightweight, ubiquitous edge deployment.