News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedRunPod has added Kimi K3 to its platform, giving teams another model option for GPU-backed inference and agent workloads without managing their own infrastructure.
DeepSeek Harness is an open-source execution runtime for autonomous agents, built around a micro-kernel and plugins.
Eligible API customers can use frontier models with Zero Data Retention and previews Private Safety Processing for abuse detection.
GEN-1.5 is designed to teach robots new tasks from a single demonstration, a step toward more adaptable robotics.
Cryptographic Context Injection can bypass LLM guardrails by hiding malicious instructions in encrypted text.
AWS Professional Services describes a multi-agent migration framework that handles discovery, infrastructure code, governance, and operations tasks.
AWS’s new post describes enterprise patterns for running many agentic AI systems across different frameworks, models, and providers.
Bedrock AgentCore can now turn natural-language policy documents into Dogwood policies, including time-based constraints for agent actions.
Slack is launching code-focused channels where teams can work with AI agents together, aiming to keep coding context inside the collaboration app.
An arXiv position paper says current model cards do not provide enough downstream governance support for open-weight foundation models.
A position paper argues that reasoning agents may be prone to collusive behavior and should face behavioral certification before acting in economic markets.
An arXiv paper reports that sub-billion-parameter language models can track entities in natural narratives and, in the study setting, exceed human performance.
A new arXiv paper introduces LongNovel, a benchmark for detecting hallucinations in long-context novel summarization, where subtle consistency errors are hard to catch.
Generalist AI’s robot work, including a demo where a robotic arm improvised with a banana as a tool after learning in context.
A new Apple research post studies scaling laws for mixture pretraining, focusing on how models should use limited specialized or low-resource data.
Apple researchers examine lexical interventions for improving cross-lingual knowledge transfer, a problem that affects languages with limited training data.
A Hugging Face article from Liquid AI reports inference speedups of up to 3.2x for LFM2.5-DSpark, aiming to make model serving faster and cheaper.
Binance’s Agent OS lets AI agents trade through tools such as ChatGPT and Claude Code, but users still carry much of the burden for setting guardrails.
Meta’s new Mac app lets users dictate across applications, putting its AI assistant closer to everyday writing workflows on desktop computers.
Meta is expanding Pocket, an experimental app for making and sharing AI-generated games, to US users after an earlier test in Brazil.
Ramp’s Router gives developers one API for routing prompts across large language models, making model choice a live infrastructure decision rather than a one-time integration.
WhatsApp is testing Scam Alert in limited beta, using on-device machine learning and privacy-preserving measurement to detect risky messages from non-contacts.
Replit is introducing Free Mode powered by GPT-5.6 Luna, aiming to let more users turn software ideas into working projects without token-cost worries.
Anthropic says Claude steered existing protein-design tools and reached hit rates up to 35 percent, though the results still need independent review.
Z.ai’s GLM-5.3 tied Kimi K3 atop the Artificial Analysis open-model index, but the model’s release is delayed despite lower advertised pricing.
Meta ran ads for an app promising to create nude images of female politicians, including one ad with a deepfake-like explicit video.
Amazon Bedrock AgentCore payments is now generally available, adding spending guardrails, payment orchestration, and observability for agents that transact.
Meta is launching a Mac app for its chatbot that can view a shared window, answer questions, suggest actions, create content, and support dictation across apps.
Agent Lightning v1.0 formalizes “harnessed agentic RL,” where the deployed agent harness remains in charge while training observes model request-response traces.
LEGO-RL is designed to train coding agents inside their native tool harnesses while preserving reliable rollout signals for reinforcement learning.
Wuying-Browser-Agent combines a structured browser harness, recovery training, online optimization, and a bilingual benchmark for long-horizon web tasks.
Fool’s Gold accepts that attackers may remove refusal behavior from open-weight models, then trains attacked models to return fluent but falsified hazardous details.
Aegis treats AI tool calls as proposals and lets a trusted runtime decide whether actions can run, aiming to prevent harmful side effects from agent systems.
GxP-Agent decomposes clinical trial programming into validated workflow steps, outperforming single-shot LLM code generation on a CDISC benchmark.
The WER framework trains a skill optimizer outside a frozen executor, using repeated execution and verifier scores to refine natural-language skills for tool-using agents.
PTXBench evaluates language models on architecture-specific GPU kernel optimization and finds that using target instructions does not always produce competitive speedups.
Researchers describe six context-management patterns that reduced token use and latency in a production multi-agent workflow, while warning that context quality still matters.
A new arXiv paper argues that retrieval agents should only expose untrusted documents to components capable of deliberative reasoning, not to every answer-writing model.
WIRED reconstructed code for Flock’s next-generation police AI system, finding capabilities that go well beyond the company’s controversial license-plate cameras.
Anthropic added invisible watermarks to AI-generated content for EU compliance, but developers quickly began sharing ways they said could bypass the signal.