Anthropic Publishes Cyber Safeguards for Fable 5 and a Jailbreak Severity Framework

On July 2, Anthropic released detailed information about the cybersecurity safeguards for the redeployed Claude Fable 5, along with a draft framework for classifying the severity of AI jailbreaks.

The "safety classifier" built into Fable 5 is an AI system that detects and blocks dangerous (or potentially dangerous) cybersecurity uses, and the disclosure includes a detailed list of which harms it is designed to prevent and which fall outside its scope.

Even more notable is the draft "AI Jailbreak Severity Framework" developed in partnership with Glasswing partners. Jailbreaks — techniques for circumventing an AI model's safeguards — vary widely in severity, from unlocking mildly undesirable behaviors to unleashing broad harmful outputs. Until now, however, there has been no unified standard for assessing that severity. With this framework, Anthropic aims to enable AI developers and governments to discuss risks on a consistent basis. The company says it intends to spur discussion across academia, industry, civil society, and government.

The Lawsuit Triggered by an AI Hallucination in a Report — The Palo Alto Networks Case

A case has been reported in which a "hallucination" (output that contradicts the facts) in an AI-generated security report escalated into a real lawsuit.

A report created with AI by Koi Security, a subsidiary of security giant Palo Alto Networks, mistakenly suggested that a startup was linked to Chinese espionage activity. Claiming business damage from this false description, the startup has filed a lawsuit against Palo Alto Networks.

This is a significant case that puts legal responsibility for the accuracy of AI-generated content to the test, and it could have implications for all companies leveraging AI going forward. The reliability of AI-generated reports and the verification processes behind them are once again being called into question in the security industry.

Goldman Sachs Releases an "AI Job Apocalypse" Report; Economists Also Sound the Alarm on AI Risks

Goldman Sachs has published a research report titled "An AI Job Apocalypse?" It quantitatively analyzes the impact of AI on the labor market and has been drawing attention on Hacker News as well.

At the same time, the WSJ (Wall Street Journal) reported that "the world's top economists are sounding the alarm on AI risks." Leading economists have expressed concerns about the potential impact of AI on economic systems, and together with the Goldman Sachs report, debate over the economic effects of AI is intensifying.

A Patch to Run DeepSeek V4 Flash with 1M-Token Context on an RTX 5090

A llama.cpp patch for running DeepSeek V4 Flash with its full 1M-token context on an RTX 5090 (on a machine with 96GB of DDR5) was announced in Reddit's LocalLLaMA community.

Previously, running at 1M context required roughly 256GB of VRAM, making it practically impossible; the cause turned out to be insufficient llama.cpp support for the DSA Lightning Indexer. A community member wrote their own CUDA kernels based on an upstream PR, achieving dramatic improvements such as:

Compute buffer: ~67GiB (OOM) → 3.2GiB

Prefill speed: 56 t/s → ~263 t/s

Practical operation at 256K context is now possible

It's a great example of the power of the open-source local LLM community, opening the door to running ultra-long-context models on consumer GPUs.

Linux Kernel Developers Debate Removing AI Attribution Tags

A new debate has arisen in the Linux kernel community over how to handle "attribution tags" for AI-generated code. According to Phoronix, kernel developers are discussing whether the tags attached to AI-generated code should be removed.

How open-source projects should position AI-generated code remains a topic of ongoing debate worldwide from the perspectives of licensing, quality control, and copyright. The outcome of this discussion in the Linux kernel — the most critical OSS project of all — could influence other projects as well.

The $1.3 Million AI Infrastructure Cargo Theft That Exposed AI's "Blind Spot"

According to The New Stack, AI infrastructure equipment worth approximately $1.3 million (about 190 million yen) was stolen while in freight transit. The incident highlighted that physical supply chain security has become a "blind spot" compared to the cybersecurity efforts the AI industry focuses on.

With expensive equipment for AI data centers becoming a target, the importance of logistics and physical security is growing, not just hardening digital security.