Anthropic has restored public and enterprise access to its powerful Fable 5 AI model after the U.S. Commerce Department lifted export controls. The relaunched model features new cybersecurity classifiers to route risky requests to older versions and is now available on Amazon Bedrock. Additionally, a new benchmark shows Fable 5 matching or beating human professionals on 16.1% of freelance remote work projects.
Sources: Ben's Bites, Digg AI, Google AI Blog, Keep Bookmarks, Latent Space, Reddit AI, The Neuron, The Rundown AI
Factory has introduced Droid Shield 2.0, an enhanced code security tool designed to prevent sensitive information leakage in autonomous engineering workflows. The system utilizes fine-tuned AI models to detect and redact secrets with fewer false positives and negatives, and the company is releasing the model weights on Hugging Face for community use.
A new open-source benchmark called Senior SWE-Bench has been launched to evaluate AI coding assistants on complex, long-horizon software engineering tasks. The benchmark specifically focuses on testing how well AI models handle realistically underspecified feature requests and vague development scenarios, marking a significant step forward in automated software engineering evaluation.
Together AI secured $800 million in funding at an $8.3 billion valuation to scale its open-model infrastructure, indicating strong investor confidence in the demand for accessible AI development platforms. This investment will further enable the company to support the growing ecosystem of open-source AI models.
OpenAI collaborated with Thrive Holdings to create Tax AI, a Codex-powered agent designed to assist with complex tax returns while preserving evidence for accountant review. The system features an improvement loop where practitioner corrections refine the AI, showcasing a practical application of AI agents in a specialized domain.
While open-source large language models have shown significant quality improvements over the past eight months for personal use, they still face major reliability hurdles in commercial deployments. This gap is highlighted by the shutdown of an AI assistant service for doctor appointments, which was retired after six months of development due to an inability to guarantee 100% correct results for third-party clients.
Developer Simon Willison revealed that the new 'shot-scraper video' command, designed for recording web application demos, was largely developed, documented, and validated by the GPT-5.5 xhigh AI agent running in Codex Desktop. The AI agent successfully defined the tool's YAML format using Pydantic, constructed complex storyboard sequences, and authored its technical documentation. This development highlights a successful pattern where AI coding agents can effectively utilize a command's '--help' output as integrated, self-contained documentation to execute tasks.
The autofz meta-fuzzer, originally accepted to USENIX Security 2023 for its innovative runtime orchestration of security fuzzers, is increasingly influencing modern AI development. Researchers are applying its control-plane framework, which dynamically allocates resources among diverse workers, to solve similar coordination and budget challenges in multi-agent LLM systems.
Amazon Bedrock has integrated OpenAI's open-weight GPT OSS models and NVIDIA's Nemotron 3 models into AWS GovCloud (US) to support sensitive government workloads. The deployment features a next-generation inference engine with a zero operator access design, ensuring strict data privacy. This expansion strengthens compliance with FedRAMP High and DoD SRG Impact Levels for U.S. government agencies.
Inscribe has introduced a new agentic AI system built on Amazon Bedrock that detects sophisticated document fraud, including deepfakes, in under 90 seconds. The system leverages a multi-model approach, utilizing Anthropic Claude Sonnet for complex reasoning, Claude Haiku and Meta Llama for cost-effective data extraction, and proprietary models on Amazon SageMaker AI for pixel-level analysis. This launch comes in response to an Inscribe report revealing a fivefold increase in AI-generated document forgeries.
Amazon SageMaker AI has introduced a managed GPU compute infrastructure to deploy and run BoltzGen, a diffusion-based generative AI model for designing protein and peptide binders. The solution automates complex infrastructure management, including instance provisioning and data pipelines, to streamline drug discovery workflows. This allows academic, biotech, and pharmaceutical researchers to focus on iterative therapeutic design while reducing compute expenses.
Google released a comprehensive update to its AI agent development stack, integrating Genkit, ADK 2.0, and cloud-local machine learning capabilities within VS Code. This update aims to streamline the creation and deployment of AI agents, offering developers more robust tools and environments.
Roon from OpenAI argues that advanced AI labs must choose between focusing on model development or pivoting to renting out compute resources. Beff Jezos, in response, indicated his immediate focus was on selling compute, highlighting a key strategic divergence in the AI industry.
A BPI report indicates that a China-linked activist group has obstructed or postponed $23.6 billion worth of US AI infrastructure projects, leading to ten data center moratoria in the US. This highlights geopolitical tensions impacting critical AI development and infrastructure expansion.
An AI-powered café in Stockholm incurred $6,000 in losses over two months because its Gemini-powered agent was susceptible to scams and inefficient inventory management, leading operators to replace it with GPT-5.5. This incident highlights the practical challenges and vulnerabilities of deploying AI agents in real-world business operations.