AI news — September 13, 2026
The 20 top stories PulseAugur surfaced that day, ranked by signal across labs, papers, and developer communities.
-
Anthropic releases Claude Fable 5.1 and Claude Code for terminal-based AI development
Anthropic has released Claude Fable 5.1 and Claude Code, a new AI stack designed for software development that moves execution out of IDEs and into the terminal. This new architecture allows for continuous reasoning over long periods, enabling proactive task execution rather tha…
-
China-linked actors used US Claude AI for military tools, intel
Chinese military researchers and tech giants have been identified using Anthropic's Claude AI model for various sensitive projects. These include developing 16 software modules for electronic warfare and air defense suppression targeting Taiwan, and drafting specifications for a…
-
Anthropic's Claude Sonnet 4.5 adds developer-focused memory tools
Anthropic has released Claude Sonnet 4.5, introducing a beta memory tool that allows agents to store and retrieve information outside the standard context window. This feature, along with context editing and longer autonomous sessions, aims to improve agent development by provid…
-
Princeton researcher proposes Recurrent Looped Transformer with unbounded temporal depth
A Princeton researcher, Yifan Zhang, has proposed a novel architecture called the Recurrent Looped Transformer (RLT). This design aims to enhance the temporal depth of decoder-only language models by carrying the complete state, including the final hidden state and layerwise att…
-
Self-hosted AI dev team AICOM automates product deployment
The developer has created a self-hosted AI development team, named AICOM, as an alternative to hosted AI app builders. Unlike simpler chat-to-code tools, AICOM utilizes thirteen specialized agents, including analysts, designers, and developers, coordinated by an AI Director to m…
-
JAX3D enables hierarchical NeRF for advanced 3D rendering and reconstruction
Researchers have developed a method for creating hierarchical Neural Radiance Fields (NeRFs) using JAX and the jax3d library. This approach enables volumetric rendering, novel-view synthesis, and 3D reconstruction. The tutorial details the process of building a synthetic dataset…
-
Anthropic releases Claude Fable 5.1 with 1M context and cost cuts
Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, which are essentially the same model with different safeguard layers. Fable 5.1 is the general availability version, while Mythos 5.1 is restricted to verified organizations for specialized tasks like cybersecurity a…
-
UK launches AI bootcamps to tackle youth unemployment
The UK government is piloting AI bootcamps in Preston, Lancashire, to address youth unemployment among individuals aged 16-24 who are not in education, employment, or training (Neets). These three-week programs aim to equip participants with AI skills, teaching them to build and…
-
LLM tool calls debugged with execution receipts
This article proposes a system for debugging Large Language Model (LLM) tool calls by treating each call as a transaction with an execution receipt. This receipt, stored within the tool adapter, contains minimal data like run ID, tool name, status, start time, duration, and a su…
-
AI agents vulnerable to pre-prompt code execution
Security vulnerabilities have been identified in AI coding agents, where malicious code can be executed before the model even processes a prompt. Research by Manifold revealed that agents could be tricked into running arbitrary code via repository-local Git configurations embedd…
-
MoE Models: Understanding the Dual Parameter Count Discrepancy
Mixture-of-Experts (MoE) models present a unique challenge in parameter counting, as they possess two distinct counts. The publicly advertised parameter count typically refers to the smaller, active set, while the true, larger total parameter count is significantly greater, ofte…
-
AI Agent Tooling Vulnerable: Only 3% of MCP Servers Validate Outputs Correctly
A new tool called mcp-drill has revealed that only 3% of 31 tested MCP servers can correctly validate tool outputs against their declared schemas. The tool, designed to test contract compliance in MCP servers, found that many servers either lack schemas entirely or use schemas t…
-
DwarfStar 4.1 AI model tested against DS4 in Sudoku challenge
The article compares the performance of two AI models, DwarfStar 4 (DS4) and DS4.1, on a Sudoku-solving challenge. It aims to demonstrate the impact of a point release on an agent's capabilities using the same set of tasks.
-
AI models learn to predict future inputs for faster inference
Researchers are exploring techniques for AI models to predict future inputs, a method that could significantly speed up inference times. This approach involves training models to anticipate the next steps in a sequence, thereby reducing the computational load during real-time pr…
-
Anthropic's Claude gains computer control, raising user autonomy questions
Anthropic's Claude AI can now interact with a user's computer, a development that raises questions about control over digital systems. This capability moves beyond simple productivity enhancements to redefine how users operate digital environments. The integration signifies a sh…
-
DotnetFastMCP v2.1.1 fixes runtime crashes and improves AI tool integration
A recent update to the DotnetFastMCP library, version 2.1.1, addresses two critical issues for .NET developers building AI-powered tools. The first fix resolves a runtime crash caused by un-registered tool classes with constructor injection, ensuring these classes are automatica…
-
Claude Fable 5.1 cracks 370-year-old cipher, solving historical cryptogram
The AI model Claude Fable 5.1 has successfully deciphered the Cyphral Distich, a 370-year-old cipher that had previously stumped human cryptographers. The model took 44 minutes and processed 176,000 tokens to solve the puzzle. Claude Fable 5.1's approach involved recognizing tha…
-
Claude Code recommends under 200 lines for CLAUDE.md files
The Claude Code documentation suggests a target of under 200 lines for CLAUDE.md files to ensure optimal instruction adherence. This guideline differs from the 4 MiB file size limit, which Claude Code loads fully, whereas longer files gradually consume more context. When CLAUDE.…
-
Iran exploited Claude AI for US Navy intel, bypassing safety guardrails
Iranian state-linked actors exploited vulnerabilities in Anthropic's Claude AI to gather intelligence for potential attacks on US Navy vessels. The operation involved a sophisticated jailbreak technique that decomposed requests into seemingly innocuous parts, which were then rea…
-
DeepSeek launches two new flash models, including vision capabilities
DeepSeek has released two new models, deepseek/deepseek-v4.1-flash and deepseek/deepseek-v4-flash-vision-exp:batch, on the OpenRouter platform. The v4.1-flash model is designed for high-volume, cost-sensitive tasks requiring lower latency, similar to Gemini Flash or GPT-4o mini.…