EU AI Act: Text Watermarking Becomes Mandatory on August 2 — What It Means for the OSS Community

A detailed analysis of the EU AI Act's text watermarking requirements has been drawing attention in Reddit's LocalLLaMA community. Starting August 2, providers of AI-generated text will be obligated to make their output machine-readable and detectable as AI-generated. The requirements reportedly follow a "two-layer structure": in addition to cryptographically signed metadata (such as C2PA), providers are also expected to implement statistical watermarking.

The rules target GPAI models classified as posing "systemic risk," which could include the Qwen 3.6 series, DeepSeek Flash, GLM, and Kimi. Simple OSS models (roughly the old Llama 7B class) are exempt, but many of today's major models are expected to fall within scope.

What makes the impact so broad is that it applies to every business serving EU citizens. The rules kick in even if just one user is an EU citizen — a tourist, say, or someone connecting via VPN. OSS tools such as LM-Studio, Ollama, llama.cpp, vLLM, and Hugging Face will need to comply as well. Violations carry fines of up to €35 million (about $40 million) or a few percent of annual revenue, whichever is higher.

The analysis also rounds up how major companies are doing: OpenAI is rolling out SynthID + C2PA, Google is expanding SynthID to images, audio, text, and video, and Adobe has already implemented C2PA hashes and Firefly watermarking. Meanwhile, OSS projects like vLLM, Ollama, and llama.cpp are described as "quite unprepared." A wave of "AI banners" — the next cookie banner — could wash over every medium: text, images, video, and audio.

EnterpriseClawBench: An Agent Benchmark Built from Real Workplace Sessions

Featured on Hugging Face Daily Papers, "EnterpriseClawBench" is an agent benchmark drawn from real enterprise workplace sessions rather than synthetic tasks. It comprises 852 reproducible tasks that involve reading heterogeneous files, making tool calls, and producing actual business deliverables.

While the underlying data is not public, the construction and evaluation protocols are a reusable contribution. Notably, even the best-performing configuration (Codex + GPT-5.5) reaches only a 0.663 success rate. EnterpriseClawBench argues that a single score hides important information, emphasizing the value of evaluating multiple dimensions: harness-model combinations, deliverable delivery, cost, execution time, and skill transfer.

As an attempt to quantify the "real-world business complexity" that existing synthetic benchmarks cannot capture, this research is likely to shake up how agent evaluation is done.

How to Push Qwen3.6-27B Q8 Past 100 t/s on Dual GPUs

A report shared in the LocalLLaMA community describes running Qwen3.6-27B Q8_0 on an RTX 5090 + RTX 3090 Ti setup and hitting 100+ t/s (up to 130 t/s). After roughly 70 t/s with the conventional layer split, switching to `--split-mode tensor` brought a dramatic improvement.

Because tensor split runs both GPUs in parallel on the same tensor, it works even when pairing GPUs from different generations. The setup combines a 70:30 `--tensor-split` configuration that gives more resources to the 5090 with speculative decoding using an MTP (Multi-Token Prediction) draft.

Under load, though, power draw tops 750W — the author warns the rig "makes for a proper space heater." As a reference case for probing the performance limits of practical local LLM inference, it's well worth a read.

NRI Secure Starts Offering a "Claude Mythos-Class" Vulnerability Assessment Service

NRI Secure Technologies has begun offering an assessment service that can detect undisclosed vulnerabilities "at a level equivalent to Claude Mythos Preview from US-based Anthropic." The head of Anthropic's Japan subsidiary has also reportedly commented on the new service.

AI-driven security assessment makes scale and depth possible that manual penetration testing alone could hardly achieve. The claim of Claude Mythos-level detection capability suggests the commercial state of AI security assessment is entering territory on par with Anthropic's frontier models.

"Status Quo Bias" Blocking AI Adoption — What OpenAI's and Anthropic's Japan Expansion Means for SIers

An analytical report from Noh-Research examines how OpenAI's Japanese subsidiary "Deploy Co" and Anthropic's newly established company could affect Japan's SIer business. It identifies the "preference for keeping the status quo" as the single biggest factor blocking generative AI adoption. With OpenAI and Anthropic moving in to directly support frontline operations, the key question is how they will approach the distinctly Japanese organizational resistance of "we want to use AI, but change is scary."

The report also highlights a Toyota Finance case, explaining why the company chose to run AI agents "alongside" its existing RPA robots rather than going all-in on agents. Rather than fully migrating to AI agents, it splits roles between AI and RPA — a pragmatic approach that keeps risk in check while advancing automation step by step.

---