Automated AI News Brief: Anthropic Launches Proactive Agent Claude Duet, LiteLLM Hits PyPI Supply Chain Crisis, OpenAI Lands on Google Cloud, and NVIDIA Unveils Kumo Relational

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 five defining trajectories across the AI landscape: the transition of agent UX from reactive dialogue to proactive background autonomy, mounting security vulnerabilities within the open-source AI supply chain, foundation models conquering structured enterprise databases, granular micro-optimizations in SLM inference economics, and the democratization of open-weight coding agents. From Anthropic's launch of Claude Duet—an assistant engineered to manage tasks autonomously while seeking human approval only when decisions matter—to a critical credential-stealing backdoor discovered in LiteLLM on PyPI, and NVIDIA's launch of Kumo Relational for direct relational database reasoning, AI is moving rapidly beyond conversational interfaces into operational infrastructure, core enterprise data, and mission-critical workflows.

Anthropic Unveils Claude Duet: A Proactive Personal AI Agent Powered by Opus 5.5

Anthropic has officially launched Claude Duet, entering research preview for Pro and Max subscribers in the United States. Departing fundamentally from conventional conversational chatbots that wait idly for user prompts, Claude Duet introduces a proactive personal agent paradigm designed to continuously manage complex responsibilities in the background.

Powered by Anthropic's flagship reasoning model Claude Opus 5.5, Duet is anchored by a core operational principle: "take on tasks of all sizes, keep working while you're away, and wait for your OK when it matters." Duet synthesizes an automated Morning Brief every day, summarizing urgent deliverables across executive calendars, email inboxes, recruitment pipelines, and contractual redlines. In the background, Duet autonomously advances multistep workflows—such as rescheduling meetings, compiling quarterly performance decks, preparing customer dispute refunds, and completing municipal licensing forms. Whenever financial transactions or critical legal adjustments are required (such as approving a $640 health permit fee), Duet proactively alerts the user with structured decision buttons, ensuring humans retain final authority without being overwhelmed by administrative drag.

Furthermore, Claude Duet natively integrates with over 6,000 enterprise connectors and plugins. It seamlessly maintains persistent memory, custom skills, and connected services inherited from the primary Claude application within a unified thread, marking a mature step forward toward persistent, autonomous, and cross-application AI assistants.

Source: Claude Duet | Your personal AI assistant

Critical Supply Chain Attack Hits LiteLLM: Malicious PyPI 1.107.0 Installs Backdoor and Hostwide Credential Stealer

In the open-source infrastructure arena, LiteLLM—an extensively adopted multi-provider gateway used by thousands of engineering organizations to route calls across leading LLMs—has suffered a severe software supply chain compromise. Security intelligence platform SafeDep published an emergency advisory revealing that every binary wheel across Linux, macOS, and Windows published for litellm==1.107.0 on PyPI contained a malicious credential harvester and remote code execution (RCE) backdoor.

Technical post-mortems show that the attackers exploited Maturin's build manifest configuration by injecting data = "cook", smuggling malicious payloads into the wheel's purelib directory. Once installed, the package deploys three dangerous components to site-packages:

1. cook.pth: Exploits Python's .pth path configuration mechanism, automatically executing cook.py whenever any Python process launches on the machine—even if the script never imports litellm.
2. cook.py: An obfuscated credential stealer that aggressively scans the filesystem for SSH private keys, AWS/GCP cloud credentials, CI/CD secrets, browser-saved passwords, local AI tool tokens, and the database keys stored in LiteLLM proxies, transmitting encrypted archives to www.litellm[.]dev.
3. sitecustomize.py: A persistent background daemon that periodically checks a remote GitHub repository every hour for updated C2 endpoints, re-executing the credential stealer and running arbitrary remote code payloads.

SafeDep verified that the official upstream GitHub repository (BerriAI/litellm) does not contain the injected code, indicating a compromised maintainer token or compromised build/release pipeline. PyPI has since purged release 1.107.0. Security teams are strongly advised to audit all development environments and build servers immediately; any host where litellm==1.107.0 was installed must be treated as compromised, requiring complete credential revocation and rotation across all cloud providers and LLM API keys.

Source: Compromised LiteLLM 1.107.0 on PyPI Installs a Credential Stealer and Backdoor

OpenAI Models and Codex Arrive on Google Cloud: Multicloud Distribution Reshapes Enterprise AI

OpenAI and Google Cloud have formally announced a landmark partnership to bring OpenAI's complete portfolio of frontier models and Codex developer workflows to Google Cloud Vertex AI's Model Garden, distributed directly through Google Cloud Marketplace.

The commercial ramifications of this agreement are substantial. Enterprise customers with pre-existing Google Cloud spend commitments can now draw down their contractual GCP balance to fund OpenAI inference workloads. While official statements reaffirm that Microsoft remains OpenAI's primary cloud infrastructure partner, the launch of OpenAI models on Google Cloud signals a decisive pivot toward a multicloud enterprise distribution model, breaking Azure's exclusive hosting monopoly over commercial OpenAI offerings.

With Anthropic already broadly accessible across AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI Foundry, OpenAI's arrival on Google Cloud illustrates how frontier model creators are embracing vendor-neutral distribution channels to capture large enterprises that demand multi-cloud resilience and flexible procurement vehicles.

Source: OpenAI models are coming to Google Cloud

NVIDIA and Stanford Researchers Introduce Kumo Relational: First Foundation Model to Beat Supervised ML on Relational Databases

While foundational language and vision models have completely transformed unstructured data handling, the operational core of modern enterprises remains anchored in relational databases (RDBMS). Traditionally, predictive tasks on relational structures—such as customer churn forecasting, fraud detection, and demand planning—have demanded weeks of arduous feature engineering, table flattening, and bespoke tuning of gradient-boosted decision trees (GBDT) like XGBoost.

NVIDIA, in collaboration with Stanford University Professor Jure Leskovec and Matthias Fey's research team, has introduced Kumo Relational: a pre-trained Relational Foundation Model (RFM) designed to operate natively on enterprise databases. Kumo Relational directly ingests relational schemas, honoring foreign keys, timestamps, and multi-table interactions without requiring feature engineering or lossy tabular flattening.

Using a specialized Predictive Query Language (PQL), Kumo Relational gathers up to 10,000 historical transactions as context, generating predictions in a single forward pass via in-context learning. Evaluated on RelBench V1—encompassing 41 distinct tasks across 15 real-world databases—Kumo Relational outperformed expert-engineered GBDTs by nearly 4 AUROC points out of the box without task-specific fine-tuning. On the SAP SALT enterprise benchmark, its zero-shot performance beat AutoGluon ensembles by 8% in Mean Reciprocal Rank (MRR). Capable of scaling to enterprise databases exceeding 500 billion rows, Kumo Relational is accessible via NVIDIA's open-source GPU-native structured-data-models library and as an enterprise NVIDIA NIM microservice.

Source: NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

Dharma AI Introduces VocabX: Vocabulary Specialization Slashes SLM Inference Costs by 77%

Fine-tuning open-source small language models (SLMs) such as Qwen or Gemma for specialized enterprise tasks is standard practice. However, research from Dharma AI reveals an overlooked cost penalty inherent in conventional deployment pipelines: the general-purpose tokenizer.

Mainstream foundation model tokenizers typically feature vocabularies exceeding 150,000 entries (such as Qwen3's 151,669 tokens) to support hundreds of natural languages and programming syntax. When deploying a 4B parameter SLM to an enterprise domain (such as an automated grading system scoring Portuguese essays into structured JSON), this expansive vocabulary introduces dual inefficiencies:

1. Structural Fragmentation: Highly repetitive domain syntax—such as JSON envelope schemas, competency keys, and standard grading terminology—gets broken down into numerous fragmented sub-tokens, forcing the model to expend unnecessary autoregressive generation steps.
2. Computational Excess: At every decoding step, the GPU is forced to compute logits across 150,000+ candidate tokens, even though more than 90% of the vocabulary is never utilized in that domain, wasting memory bandwidth and FLOPs.

Dharma AI's VocabX framework reconstructs token vocabularies around the marginal computational cost of token generation rather than naive corpus frequency. Without altering the underlying model architecture or fine-tuning hyperparameters, VocabX reduced measured production inference costs by 77%, providing an actionable blueprint for high-efficiency enterprise model serving.

Source: How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications

Recursive Releases Coding Agent Model Flint (40B / 2.5B Active) Alongside MemT-1.0 Memory-Native Architecture

The open-source community witnessed significant breakthroughs today in agent-centric model design and long-horizon memory architectures with the simultaneous debut of Recursive's Flint coding model and the MemT-1.0 memory framework.

Recursive Flint is a sparse Mixture-of-Experts (MoE) foundation model explicitly optimized for coding agents and command-line execution. While housing 40 billion total parameters, it activates only 2.5 billion parameters per token. Featuring a 3:1 Sliding Window Attention (SWA) to full attention ratio and a 262k token context window, Flint achieved top performance on TerminalBench 2.1 and SWE-bench Pro v2, outperforming previous community favorites such as Qwen3.6-35B-A3B while running fluidly on consumer-grade hardware (such as single RTX 3090/4090 GPUs).

In parallel, researchers unveiled MemT-1.0 (Towards Memory-Native Evolution), addressing the severe KV cache bloating and degradation that afflict long-horizon agents. Built upon Gated DeltaNet-2 and Mixture-of-Storage, MemT-1.0 internalizes memory updates into the model's fundamental forward computation. At a 128k token context horizon, MemT-1.0-Nano achieves a 1.86x end-to-end inference speedup and a dramatic 66x reduction in persistent storage overhead compared to standard KV caching, providing an essential foundation for scalable, autonomous long-term agents.

Source: Reddit LocalLLaMA: Recursive Flint - the "Qwen3.8 35B" we've all been waiting for? / [Paper] MemT-1.0: Towards Memory-Native Evolution

IBM Debuts Granite Speech 5.0 Suite: Advancing Multilingual Speech Recognition, Edge CTC, and Diarization

IBM has open-sourced Granite Speech 5.0 on Hugging Face, releasing a comprehensive enterprise speech toolkit comprising a new 2B speech-language model and two upgraded lightweight architectures designed for production telephony and meeting transcription:

1. granite-speech-5.0-2b: A flagship speech-language model featuring robust automatic speech recognition (ASR) resilient to ambient noise and reverberation, language-ID prompting, 2,000-term keyword biasing, and enhanced long-form cross-lingual translation.
2. granite-speech-5.0-470m-turboctc: A non-autoregressive CTC model optimized for sub-millisecond edge transcription, now supporting multilingual recognition with native punctuation and capitalization recovery.
3. granite-speech-4.1-2b-plus: A specialized model engineered for speaker attribution and diarization, delivering high-precision tracking of who spoke, when, and what was said across complex multi-speaker recordings.

Source: Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster

Today's Takeaway

Today's developments underscore a decisive shift in artificial intelligence from isolated benchmark victories to systemic enterprise integration:

1. The Proactive Agent Transition: Claude Duet highlights the evolution of user interaction. Instead of waiting for prompts, agents now operate persistently in the background, filtering noise, drafting execution plans, and prompting users exclusively for high-leverage decision checkpoints.
2. Fragility in the AI Gateway Layer: The LiteLLM PyPI compromise serves as a wake-up call for AI engineering teams. As ecosystems grow increasingly dependent on intermediate abstraction libraries, hardening release pipelines and enforcing strict package signing become indispensable pillars of AI DevSecOps.
3. Structured Data Enters the Foundation Era: NVIDIA's Kumo Relational demonstrates that foundation models can ingest multi-table enterprise relational databases directly. Eliminating manual tabular feature engineering opens a transformative avenue for zero-shot business intelligence and autonomous database reasoning.
4. Systems-Level Efficiency at the Micro-Scale: From Dharma AI's VocabX cutting token costs by 77% to Recursive Flint's 2.5B active parameter MoE and MemT-1.0's 66x memory compression, hardware utilization and cost discipline are driving open-weight models closer to widespread enterprise viability.