Overview
10 stories in this issue. The first 3 are today's priorities.
Hot model updates
- Top · Nous Research open-sources Arete-27B, cutting Qwen3.8's propaganda rate from 22% to under 1%
- Top · DeepSeek V4.1 Flash completes 72.8% of attack objectives on Vals CyberOffense at about $3 per run
- Top · Auditing agent logs, GPT-5 catches about 98% of reward hacking but raises many false alarms
Global AI news
- Apiiro finds secrets in 1,122 public Claude Code settings files
- Anthropic tells customers an unnamed security firm accessed its internal systems without authorization
- NYT: a Chinese firm cleared by the US to receive Nvidia H200 chips has an empty office
- Physicl raises a $10M seed round to turn e-commerce 3D into robot training data
- Open-source agent Observer uses local models to watch your screen, camera and audio, and alerts you when something happens
Regional and early signals
- GitHub's new Git storage architecture, Rubicon, handled 5,000 concurrent agents on one repository in internal tests
- GPTZero says an Accenture report contains 15 fabricated citations; Accenture blames human error

Jiufeng graphic based on the sources cited in this issue.
Hot model updates
01/10
Nous Research open-sources Arete-27B, cutting Qwen3.8's propaganda rate from 22% to under 1%
Nous Research ran supervised fine-tuning on Alibaba's Qwen3.8-27B with 53 million tokens to produce Arete-27B, and released the weights and training data under Apache 2.0.
Nous Research is best known for its open-source agent Hermes. It fine-tuned Alibaba's Qwen3.8-27B to remove Chinese state propaganda from the model's responses. Nous pitches the result, Arete-27B, as offering "Western values" in place of Beijing's party line.
| Model | Propaganda rate (%) |
|---|---|
| Qwen3.8-27B (original) | 22 |
| Arete-27B (fine-tuned) | <1 |
- Base model: Alibaba Qwen3.8-27B
- Training: supervised fine-tuning on 53 million tokens
- Released: model weights, test dataset and training data
- License: Apache 2.0, published on Hugging Face
Nous says this shows the political bias in Chinese open-weight models is only "a thin layer". The Decoder also reports that Nous is using its funding to move into enterprise sales with "Hermes for Businesses".
Limitations: The propaganda rates and the claim that general capabilities stayed intact both come from Nous's own evaluation. The "thin layer" conclusion rests on this one model; the source does not show whether it holds for other Chinese models.

Image source: huggingface.
Source: Original source Hugging Face: NousResearch/Arete-27B · The Decoder
02/10
DeepSeek V4.1 Flash completes 72.8% of attack objectives on Vals CyberOffense at about $3 per run
On a benchmark that simulates corporate networks, the open-weight DeepSeek V4.1 Flash 1017 finished 8 points behind OpenAI's restricted GPT-6 Sol.
Vals CyberOffense builds realistic simulated corporate networks. Each agent works alone from an isolated Kali Linux attack machine. It starts with only a public web address and gets no human help. To succeed, it has to move from the web application through Windows servers to admin privileges. Every action is logged externally and scored automatically.
| Model | Attack objectives completed (%) | Cost per run (USD) |
|---|---|---|
| GPT-6 Sol (Daybreak) | 80.8 | ~22 |
| DeepSeek V4.1 Flash 1017 | 72.8 | ~3 |
GPT-6 Sol is part of OpenAI's restricted Daybreak cyber program. Vals researchers note that once a model's weights are public, no provider can control or restrict how it is used.
Limitations: These are results in simulated networks, not real-world attack success rates. The two models also have very different access: GPT-6 Sol is limited to the Daybreak program, while DeepSeek V4.1 Flash can be downloaded freely.
Source: Single report The Decoder
03/10
Auditing agent logs, GPT-5 catches about 98% of reward hacking but raises many false alarms
Andy Hall, a researcher at the Anthropic Institute, replicated and extended a METR study. His conclusion: AI auditors are good for estimating how often agents misbehave, but not for safety checks that must catch every case.
METR's MALT dataset holds logs of frontier models working through HCAST and RE-Bench tasks. Human reviewers checked part of these logs for misbehavior in October 2025. Hall tested five models on the same logs: GPT-5, GPT-5 nano, GPT-6 Astra, Opus 4.8 and Gemini 3.8 Flash.
GPT-5 found about 98% of reward-hacking cases. But the harder the models tried to catch every case, the more harmless actions they flagged as suspicious. That makes them better suited to measuring how often things go wrong than to checks that must reliably catch every incident.
Limitations: The analysis was published on Hall's personal Substack, "Free Systems". Hall is on leave from Stanford and says the post reflects only his personal views.
Source: Single report The Decoder
Global AI news
Claude Code saves commands that developers approve to .claude/settings.local.json inside the project folder, and passwords and API keys in those commands have ended up in public GitHub repositories.
Security company Apiiro published the research on November 5. Researchers Shmulik Cohen and Matan Giladi found 1,122 public copies of the file containing at least one secret. Broader scans turned up 2,539 secrets across 1,994 public repositories.
RuntimeWire advises Claude Code users to rotate exposed secrets, scan agent folders and Git history, and replace commands that contain literal credentials with patterns and secret-manager variables.
Limitations: All counts come from Apiiro's own scans. The source does not say whether Anthropic will change how the file is written by default.
Source: Original source The Cyber News on X · RuntimeWire
05/10
Anthropic tells customers an unnamed security firm accessed its internal systems without authorization
The notice says Anthropic cannot confirm whose information was viewed, yet also says recipients face no heightened risk.
According to the notice as quoted by Jamieson O'Reilly on X, Anthropic told customers that a cybersecurity firm accessed its internal systems without authorization. The notice does not name the firm or say which systems were affected.
RuntimeWire notes that with so little detail, customers cannot judge their own exposure. It is also unclear whether this was bug-bounty research that went beyond its scope or access that was never authorized at all.
Limitations: The notice is only available through O'Reilly's quote. Anthropic has not publicly named the firm or the affected systems.
Source: Original source Jamieson O'Reilly on X · RuntimeWire
06/10
NYT: a Chinese firm cleared by the US to receive Nvidia H200 chips has an empty office
A New York Times visit found that New Matrix, which is on the list of firms approved to handle Nvidia's H200 chips, lists an address that is an empty shell.
The Verge, citing The New York Times, reports that New Matrix's office is "a hollow gray shell with wires dangling from the ceiling where lights should be." The company was created 17 months ago, has little presence online, and is missing from the building's company directory.
According to the report, the US government is approving sales of Nvidia's AI chips to Chinese firms with ties to blacklisted companies. This has raised questions about whether the Commerce Department and Nvidia are doing proper due diligence.
Limitations: This is the conclusion of a media investigation. The evidence on whether New Matrix has real operations is limited to the empty office and the lack of information online.
Source: Single report The Verge
07/10
Physicl raises a $10M seed round to turn e-commerce 3D into robot training data
Founder Alex de Vigan is repurposing the 3D pipeline behind nfinite to generate labeled synthetic scenes for world models and robots.
Physicl announced a $10 million seed round from Inovia Capital on November 6, 2026. It generates scene variations under controlled settings, each with depth, segmentation and camera data attached. The aim is to cover rare failure cases that real video seldom captures.
Limitations: RuntimeWire notes that the approach only works if the synthetic worlds transfer to real-world physics. The report gives no data on how well they transfer.
Source: Multiple outlets RuntimeWire · TechCrunch
08/10
Open-source agent Observer uses local models to watch your screen, camera and audio, and alerts you when something happens
Observer lets users define "when X, then Y" monitoring tasks, runs entirely on local models, and its developer says data stays 100% private.
Each Observer micro-agent runs in a Sensors → Model → Code loop. The project's examples include calling you when a training run crashes, sending a Telegram when an item is back in stock, and sending a WhatsApp when an appointment opens up.
- Web app: no install; open app.observer-ai.com and download a model to run in the browser
- Desktop app: macOS, Windows and Linux; ships with llama.cpp and runs any GGUF+mmproj model
- Bring your own server: point it at any OpenAI-compatible endpoint, such as Ollama, llama.cpp or vLLM
Limitations: These features are described by the developer in the repository. No independent evaluation was available.
Source: Original source GitHub: Roy3838/Observer
Regional and early signals
09/10
GitHub's new Git storage architecture, Rubicon, handled 5,000 concurrent agents on one repository in internal tests
GitHub says that in internal stress tests, Rubicon let a single repository host 5,000 agents working at once and absorb 7,200 pushes per second.
According to IT Home, GitHub unveiled the "Rubicon" storage architecture at GitHub Universe and described it in a blog post on November 5. It can handle thousands of writes per second and scale reads horizontally on demand. Today, even GitHub's busiest repositories peak at only tens of writes per second. GitHub says daily requests to one typical repository grew about 4x between March and May.
Limitations: The concurrency figures come from GitHub's internal stress tests. Rubicon is in a limited private preview, with wider availability planned for early next year.
Source: Single report IT Home (Chinese-language source)
10/10
GPTZero says an Accenture report contains 15 fabricated citations; Accenture blames human error
An Accenture report on AI in African agriculture, written with Madrid's IE University, allegedly includes fabricated case studies. Accenture denies the errors came from AI hallucination and is revising the report before republishing it.
According to IT Home, GPTZero's analysis found that Accenture's "Leapfrogging Development" report may contain 15 fake citations and fully fabricated case studies. The report examines how AI can raise agricultural productivity in Africa.
An Accenture spokesperson said the company's review shows the errors came from human error, and that there is no evidence they were caused by AI hallucination. The company is correcting the report and will republish it once the work is done.
Limitations: The two sides disagree on whether AI caused the errors.
Source: Single report IT Home (Chinese-language source)
This digest is compiled by AI from the sources linked in each item and published after automated fact-checking, without article-by-article human review before publication. If anything differs from the original, the original source prevails. How this digest is made, or report an error: About this digest
Create AI images and videos with sound in your browser: write a prompt, pick a model, download the result. 5 free credits on sign-up, no credit card.
Start creating free

