Amazon SageMaker AI has launched a new 'KV_AWARE' routing strategy that directs LLM inference requests to instances based on their actual KV cache state rather than just prompt headers. This optimizes multi-turn conversations and long-context workloads by ensuring requests land where relevant data is already cached.
CISA has issued an urgent advisory regarding CVE-2026-114938, a critical memory-corruption vulnerability in the x86-64 HotSpot JVM. This flaw allows unauthenticated attackers to execute remote code via maliciously crafted Base64 input.
Microsoft has introduced 'Perception', an agentic security platform designed to manage the risks posed by AI-driven, automated cyberattacks. The platform coordinates red, blue, and green team agents through automated playbooks to detect, investigate, and remediate vulnerabilities at machine speed. This release also includes new capabilities for securing 'agentic insiders'—AI agents that act as autonomous entities within corporate networks.
Microsoft and Salesforce have launched Copilotforce, an initiative that integrates Salesforce customer data and workflows into Microsoft Copilot and Teams. The partnership uses authenticated MCP access to allow AI agents to act on CRM data while respecting existing security and permissions.
AWS introduced the Social Messaging Agent Runtime Toolkit (SMART), an open-source SDK and Terraform stack designed to manage the complexities of WhatsApp integrations for AI agents built on Amazon Bedrock AgentCore. It handles authentication, media processing, conversation memory, and 24-hour message window constraints automatically.
OpenAI has launched a public beta for the Responses API that allows you to select specific background processing windows. By opting for a two-hour or eight-hour completion window, your company can achieve substantial cost savings on token usage.
NVIDIA has released a technical primer on co-designing sparse attention for long-context models. By focusing on GPU-friendly indexer kernels—which identify relevant tokens before full attention—developers can significantly improve inference efficiency for context lengths surpassing 100K tokens.
Anthropic has launched a beta feature called Linked billing, enabling end-users to cover API costs directly through their own Anthropic organizations. By utilizing the user-profiles-billing-2026-10-01 beta header, developers can now toggle payment responsibility per request.
Amazon has published a three-phase roadmap to help platform and application teams operationalize Responsible AI for agentic systems. It addresses the challenges of nondeterministic behavior, data boundary isolation, and regulatory compliance as agents move from pilot to production.
Google has launched 'Tasks in Gemini,' a new capability allowing the assistant to perform background, multi-step actions. The update includes features for automated purchasing via virtual Google Pay cards, task management via SMS, email-based workflow automation, and 'Agentic Calling' for making phone inquiries and bookings.
Microsoft security researchers have simulated a scenario where attackers use phishing to compromise AI coding tools via OAuth device code flows. By tricking users into approving a legitimate-looking device request, attackers could gain access to private repositories, session history, and potentially control active command-line agent sessions.
NVIDIA unveiled 'Nemotron for Science', a toolkit combining the Nemotron 3.5 Super VL model with specialized chemistry datasets and post-training recipes. This platform allows researchers to deploy local agents that can autonomously plan experiments, execute code, and validate data findings.
The NCSC warns that standard security monitoring is no longer enough to catch sophisticated attackers, especially as adversaries adopt AI to increase the speed and scale of operations. Organizations must shift toward proactive threat hunting, focusing on enduring attacker behaviors (TTPs) rather than just changing indicators of compromise.
Siemens has patched two security vulnerabilities in the Mendix Studio Pro MCP Server. These flaws could permit remote attackers to interact with the server or execute unauthorized actions via browser-based CSRF attacks.
Researchers introduced OmniTag-Bench, a new five-language framework for evaluating multimodal LLMs on simultaneous speech transcription and metadata tagging. Instead of using separate specialist models, the study demonstrates that a single, fine-tuned multimodal LLM can extract transcripts, emotion, gender, language, and named entities in a single inference pass.
Researchers have introduced Trident, a framework designed to improve LLM-based agents by solving three information bottlenecks: ineffective memory retrieval, redundant behavioral candidates, and risky whole-harness rewriting. By implementing boundary-aware memory reading, diversity-aware routing, and order-aware patching, Trident consistently outperforms existing reflection and memory-based methods.
Researchers have introduced EndoUT, an adversarial framework that identifies universal prompt triggers capable of significantly degrading LLM utility. By analyzing how a model's internal response-onset preferences are disrupted, attackers can create fixed strings that force performance drops across various benign tasks without needing target labels or feedback.
Amazon Bedrock has launched a new memory-aware capability for AgenticRetrieveStream. This feature automates the storage, extraction, and retrieval of user-specific facts and conversation histories, allowing agents to provide context-aware responses while maintaining strict document-level access control.
AssemblyAI has launched STT-3.7, a new speech-to-text model replacing the Universal family. It offers increased language coverage and significant improvements to speaker identification accuracy.
AssemblyAI has introduced AssemblyAI-STT-3.7-Realtime-Agent, a speech-to-text model specifically designed for voice agent interaction. It aims to solve common issues in live telephony, such as distinguishing user speech from background noise and capturing critical data points like numbers at the end of a sentence.
AssemblyAI has launched the STT-3.7 model family, featuring dedicated versions for pre-recorded audio, live conversations, and voice agents. Additionally, the company introduced its first text-to-speech engine, AssemblyAI-TTS-1, optimized for high-speed, natural-sounding interactions.
Anthropic has implemented a text-based watermark for its newest Claude models, including Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5. The watermark applies when using default sampling configurations but may be absent if users modify temperature, top_p, or top_k settings.
MoEngage has integrated AI agents into its customer engagement platform by using the Model Context Protocol (MCP) and Amazon Bedrock AgentCore to expose existing APIs as agent-callable tools. By treating its platform as an architecture of isolated sessions and standardized tool schemas, the company enabled teams to automate workflows and decision-making within a single conversation.
TokenEconomizer is a new open-source plugin that allows AI coding agents to learn from past sessions within a specific repository. By tracking recurring patterns, pruning unnecessary tools, and preventing repetitive failure loops, it significantly reduces token consumption without altering the agent or its base model.