… OpenAI has also disclosed the identified zero-day vulnerability in the internally-hosted third-party software to the vendor for a fix. To aid Hugging Face in handling similar AI agent attacks in the future, OpenAI has included them in its trusted access program to improve their defenses. …
… In the same vein, Microsoft unveiled the following technologies too: Preview available: Microsoft Execution Container MXC SDK for OS-level control over agent execution Through Defender AI model scanning, developers can inspect model artifacts Generally available: Windows 365 for Agents, generally a… …
… In benchmarks measuring coding, agentic, and general capabilities, Qwen3.8-Max beats Fable 5 and GPT-5.6 Sol across 7 different evaluations. In multimodal and vision benchmarks, it leads both models across 36 different tests. Find the full list of benchmarks below. …
… However, for most advanced tasks, including coding, Opus 4.8 still remains the superior option that’s available to the general public. Anthropic says Sonnet 5 shows a lower rate of undesirable behaviors than Sonnet 4.6 and is generally safer to use in agentic contexts. …
… Generally available, starting today. 🧵 GoogleIO pic.twitter.com/jLhqozutwG — Google @Google May 19, 2026 Here's how Gemini 3.5 Flash compares with Gemini 3.1 Pro and other models from competitors : Benchmark Gemini 3.5 Flash Gemini 3 Flash Gemini 3.1 Pro Claude Sonnet 4.6 Claude Opus 4.7 GPT-5.5 Te… …
… Instead of relying on a single AI model for vulnerability scanning, this new system uses more than 100 specialized AI agents across different frontier and distilled models. …
… In some of the aforementioned benchmarks, the models were tested using different agent harnesses, making direct comparisons less straightforward. Kimi K3 is now available through Kimi’s web and desktop apps, Kimi Code, and its API. …
… First, Microsoft announced MAI-Cyber-1-Flash, its first in-house AI model built specifically for cybersecurity. Second, the company announced Project Perception, an agentic AI system that uses specialized AI agents to reason over security data, tools, and workflows to deliver continuous defense. …
… It can also act as a main agent that creates a plan and delegates work to parallel subagents. The new Muse Spark 1.1 supports a 1-million-token context window, a significant upgrade from Meta's earlier models. Meta claims that this model can actively manage that context during long workflows. …
… According to Meta, Muse Code is the best harness for its model Muse Spark 1.2, as they were trained together. The agent includes several background subagents that remain active throughout a session. …
