Security Standard SAFE for Agentic AI Launches with 120+ Organizations
To coincide with the opening of the Black Hat conference, the Open Secure AI Alliance — whose members include NVIDIA, Cisco, CrowdStrike, Hugging Face, and Red Hat — has announced a proposal for new guidelines to strengthen the cybersecurity of agentic AI. The coalition, which now counts more than 120 member organizations, has joined forces with the Linux Foundation to open a public comment period for the guidelines, called SAFE (Shared AI Findings Exchange).
SAFE aims to confidentially collect and analyze incidents and near-misses involving agentic AI, notify affected parties, identify recurring control failures, and then circulate evidence-based operational recommendations back through the entire ecosystem. The coalition's position is that if defenders are to move at agent speed, a trusted ecosystem sharing threat intelligence is the best approach — a framework designed to turn individual companies' incidents into collective defensive strength.
What draws attention is that, alongside the guidelines, member companies released open source tools all at once. The shared premise is that an AI agent is not just a model but a "system" composed of identity controls, harnesses, guardrails, logging, and evaluation — one that vulnerability scanning alone cannot fully protect.
NVIDIA released NOOA, a research harness that makes it easier to test and audit agent behavior, along with OpenShell, a runtime that restricts what agents can touch, and Garak, a vulnerability scanner for LLMs, among others. Amazon, joining the alliance at the same time, open-sourced the agent-building toolkit Strands Agents and the authorization language Cedar; Microsoft's AI Red Team offered the automation tools PyRIT, RAMPART, Clarity, and Assert; Cloudflare contributed a vulnerability-discovery harness; and Capital One provided VulnHunter. Uber open-sourced core components of ADR, its agent detection and response system that processes 200,000 sessions per day. On the identity layer, Okta, Palo Alto Networks, and Red Hat are contributing, while LangChain is contributing on the recovery layer with fallback functionality. The idea that "security is strongest in layers" is taking shape as the defensive stack for the agent era.
DeepSeek-V4-Flash-0731 Reaches 1M-Token Context Locally
How far can you push DeepSeek-V4-Flash-0731, released at the end of last month, on hardware you already own? Community validation has moved another step forward. According to a report on LocalLLaMA, a desktop pairing an RTX 5090 (32GB) with 256GB of DDR5 memory achieved a full run at a context length of 1 million (1M) tokens using vLLM's CPU/RAM offloading. It recorded roughly 800 tokens per second at prefill and more than 15 tokens per second at decode — a level considered solid enough for coding in agentic use.
The reporter measured that the acceptance rate of speculative decoding (DSPark) affects throughput almost one-to-one, and suggests that dynamically switching draft depth between reasoning and normal generation is likely to be "the biggest remaining optimization."
Around the same time, QuarkStar, a small engine inspired by Antirez's DwarfStar, was also released. It natively supports Vulkan (Linux) and Metal (Apple Silicon), and can run Qwen3.6-35B-A3B entirely in memory on a 16GB setup. On the developer's AMD BC-250 (a roughly $150 secondhand board with 16GB of integrated GDDR6), it reportedly achieved about 82 tokens per second at decode with a 2K context. The road to approaching frontier-class performance without thousands of dollars of dedicated hardware is being paved, little by little.
In the Japanese-speaking community, a proof-of-concept repository has also been published with a configuration and patches for running DeepSeek-V4-Flash-0731 in production on a single AMD Instinct MI300X, sparking discussion of FP8 porting pitfalls and the economics of on-premises operation.
Apple v. OpenAI: Trade-Secret Investigation Widens
Apple is expanding its trade-secret investigation involving OpenAI. In new court filings, Apple indicated that additional former employees may have retained or accessed confidential information. The pool of individuals implicated is reportedly wider than initially thought, and at this point neither the other side's rebuttal nor the final outcome is settled. It is still at the stage of watching how the litigation unfolds.
Blackstone Sounds Out $36 Billion-Plus Debt Financing for Anthropic
Major financing to fund Anthropic's procurement of Google chips is moving behind the scenes. In early soundings Blackstone held with investors, a debt package of at least $36 billion was contemplated. The size, the structure — and even whether Blackstone ultimately leads the deal — are still under discussion and subject to change. Anthropic was also recently reported to have a roughly $10 billion contract with Volta, an AI cloud upstart, one in a string of financing moves aimed at securing compute resources.
SK hynix and SanDisk Announce HBF, a New Memory Spec for AI Inference
In collaboration with SanDisk, SK hynix has announced HBF (High Bandwidth Flash), a new memory spec aimed at eliminating AI inference bottlenecks. It is expected to deliver bandwidth of up to 3TB/s. Memory bandwidth tends to be the limiting factor when running inference on large models, and flash-based high-bandwidth memory could reshape the balance between cost and capacity. That said, when it might reach consumer local environments — and at what price — remains unclear, so for the time being it looks positioned for data centers.