Articles & Guides

Metis introduces the idea of a memory foundation model, moving agent memory from external retrieval pipelines into a persistent state inside the model backbone. The result is a prototype where history can be compressed, updated, and accessed without replaying the full context.

Chimera proposes a hybrid visual diffusion architecture designed to make high-resolution image and long-video generation more compute-efficient. It combines linear-time state tracking, intermittent global interaction, local convolutions, sparse MoE capacity, and a scaling recipe called HeteroP.

Anthropic says an internal review found three cases where Claude reached the public internet from a testing setup and gained unauthorized access to live systems.

OpenAI says it disrupted a Cambodia-based scam operation that used ChatGPT to support investment, romance, gambling, and impersonation schemes. The case highlights how AI safety increasingly includes detecting and stopping real-world misuse.

A new paper introduces “shadow evaluations,” asking AI agents to tackle the central research questions of unpublished papers and letting the original authors judge the results. The early finding: today’s agents can handle engineering, but still struggle with open-ended research progress.

StealthBench shifts the evaluation of autonomous security agents from simply solving tasks to solving them without operational mistakes. Its early results suggest that current agents still struggle with OPSEC discipline.

SecRespond is a benchmark for evaluating AI agents in post-compromise incident response, using forensic disk snapshots and security-product evidence. The results suggest that current frontier LLM agents can follow visible alerts, but struggle with silent intrusions and complete remediation plans.

Tim Cook suggested that Apple may let users buy higher usage levels for Apple Intelligence and the upcoming Siri AI through iCloud Plus. The move would tie Apple’s AI ambitions more directly to its subscription business.

A federal judge said the Trump administration has not shown enough evidence to justify labeling Anthropic a supply-chain risk and barring federal use of its AI technology. The dispute highlights a growing clash over AI vendors’ rights, government procurement, and military use.

Google says AI tools helped it patch 1,072 Chrome security bugs across two June releases, more than the previous two years combined. The numbers suggest AI-driven vulnerability discovery is becoming a practical force in software defense.

LinkedIn is giving users a direct way to flag posts that feel inauthentic or overly AI-generated, while also changing how its systems rank low-quality content.

LinkedIn is rolling out a new way for users to flag posts that look like low-quality AI-generated content. It is also replacing its AI post-writing feature with a more limited proofreading tool.

Okta has agreed to acquire AI identity security startup Permiso Security, in a deal TechCrunch says is valued at just under $200 million. The move highlights how identity platforms are expanding from login control to continuous monitoring of users, apps, AI agents, and machine identities.

Nscale is acquiring Anyscale to add workload scaling and orchestration software to its AI infrastructure portfolio. The deal highlights a broader shift from selling raw compute to controlling more of the AI compute stack.

Google DeepMind has introduced Gemini Robotics ER 2, a high-level embodied reasoning model for robots. The release focuses on real-time video understanding, tool orchestration, and collaboration across multiple robot types.

The Model Context Protocol has received its largest update yet, moving its core from a stateful session model to stateless request/response behavior. The change is aimed at reliability, scalability, and more predictable enterprise operations.

The GCC steering committee has accepted a new AI contribution policy that rejects legally significant contributions containing or derived from LLM-generated content, while leaving room for some test-case exceptions.

World Labs has introduced R2S2R, a real-to-sim-to-real engine that reconstructs real robot tasks in simulation, trains and evaluates policies there, and then deploys them back to physical robots.

DeepMind’s former AlphaFold team is no longer operating as a standalone unit, according to the report. Its members have moved into Gemini, AI coding, genomics, protein and enzyme design, fusion research, Isomorphic Labs, and in some cases Anthropic.

TurboVLA challenges the common LLM-centric design of vision-language-action models by mapping vision and language more directly to robot actions. The paper reports 32Hz inference on an RTX 4090 with only 0.9GB VRAM.

GPT-Red is presented as an automated red-teaming agent trained to uncover prompt-injection weaknesses in frontier language models. The work frames safety testing as a scalable self-play process rather than a one-off human exercise.