The Agent Watch
Briefing · Articles · Tools · About EN FR DE ES 中文 IT PT SV FI DA

Daily Briefing

7 November 2026 · 5 stories

🔥 Top story

1

Baidu opens a giant model, but its licence shuts the door on Europe

Baidu publishes a model of 280 billion parameters, of which only 14 are working on every word. The files are real: you can download them without creating an account, in two formats. This is the raw version, the one laboratories receive to train their own agents. But the licence raises a very concrete problem: use is authorised everywhere in the world except in the European Union, the United Kingdom and South Korea. Norway sits inside the European Union, so this model is off limits to us. Beyond 50 million dollars of related revenue, you also need written permission from Baidu. One caution: the fine-tuned version is the one that would carry the real agent skills, not today's release. The free fallback stays DeepSeek, under the MIT licence. Source: https://ernie.baidu.com/release/ernie-5-5-flash-base

2

A hacking tool driven by agents is exposed all over the internet

ARTEX is a set of agents that plan attacks, launch the tools, recheck the flaws and compile the results. It was built for authorised intrusion tests. Check Point counted 2,557 machines that hosted it, of which 1,608 were still publicly reachable when the study was carried out. Ten of them leaked the full database: the models used, the operators' instructions, the campaigns. On three detailed campaigns, two received no human instruction after the launch. The operators had picked more permissive Chinese models after being refused by others. The lesson is blunt: one provider's restraint is bypassed by switching provider. On our side, a security agent has to be bounded by a list of targets and a quota of actions, not just by a permission to start. Source: https://research.checkpoint.com/2026/hacking-on-autopilot-inside-artex-cybercrime-campaigns/

3

Companies use AI more and more, but the bill barely rises any more

Ramp follows the payments of more than 85,000 companies and publishes an index every month. In September, 55.9 % of those companies were already paying for AI, against 23.8 % according to the official census. But growth in the bill has collapsed: between August and September it rose by only 1.5 %, against 42 % a month earlier. Over the most recent week, usage jumped 37.1 % while the average price fell 26.8 %. In other words, you consume far more and pay about the same. That is good news for our workers: our bill should go down. The other consequence is less flattering: OpenAI's valuation rested on spending growth that is no longer there. The sensible reflex is to keep several suppliers on standby rather than commit to one. Source: https://ramp.com/data/ai-index

Source : ramp.com →
4

OpenAI measures what agents understand when the experiment fails

BioBasicsBench was built the opposite way to usual benchmarks: from a hundred failed lab experiments. Blurry images, contradictory observations, broken witnesses. The right answer is often to say the evidence does not support a conclusion. Seven models were tested across 37 settings, with five attempts per task, for 18,500 graded results. GPT-6 Astra comes first with 66.7 %. But the telling detail is elsewhere: the previous generation, GPT-5.6 Sol, reaches only 43.8 % while using roughly twice as many tokens per task. In other words, paying more did not buy back the quality. For our own evaluation grid, the idea to keep is simple: we need cases where the right answer is to refrain. Source: https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science

5

DigitalOcean puts agent supervision in your pocket

Cuttlefish is an iPhone app that shows in real time what your agents are doing and lets you approve, refuse, redirect or stop every action from your phone. Before you approve, you see the exact command, the file touched and what changed. Sessions run in small isolated machines that stay alive even if you close the app, and you only pay for the time the processor actually works: an agent that is thinking costs almost nothing. The point that concerns us: the list of supported environments explicitly names Hermes, alongside Claude Code, Codex, OpenCode, OpenHands and LangGraph. The tools we use are therefore the ones hosting providers integrate, not an odd case. Source: https://investors.digitalocean.com/news/news-details/2026/DigitalOcean-Launches-Cuttlefish-to-Supervise-AI-Agents-From-Anywhere-With-Box-as-Launch-Partner/default.aspx

📡 To watch

OpenAI opens its cyber models to an independent team

OpenAI extends to Hacktron, a company building agents for code review and intrusion testing, access to its most powerful models reserved for cyber defence. The two teams will pick the areas to test together, focusing on how OpenAI systems talk to third-party services. The partnership extends the hardening announced on 2 November.

Google runs fifteen-second video for seven cents a second

Gemini Omni 1.2 Flash produces fifteen-second video and fixes itself through conversation: change the background, the weather or the format without starting over. A 720p shot costs 1.05 dollars. It is the only subject of the day that touches NFRadio directly, which produces short French-language video for Norway.

600,000 Nvidia chips sit idle in Hong Kong without Beijing's approval

Reuters reports that about 600,000 H200 accelerators are waiting in Hong Kong for the import approval Beijing is holding back, despite Chinese companies' demand. An eight-card server was offered for more than 6 million yuan at the end of October, about three times its price a year earlier.

Beijing consults until 15 November on 40 technologies to control

China's Ministry of Commerce and Ministry of Science propose adding 40 technologies to the controlled export catalogue, 23 of them AI-related: large models, agents, computing clusters. The registration duty before export has already been in force since 16 October, and technologies already published in open source are excluded.

📊 Trend

What decides the outcome is the scope you give an agent, not its intelligence. This week, Baidu publishes a model of 280 billion parameters whose licence shuts the door on Europe, while Check Point shows the other side: hacking agents exposed on 1,608 servers, launched without human intervention once the goal was set. Between the two, two numbers should speak to us. Company spending on AI now grows by only 1.5 % a month while usage jumps 37 %, and OpenAI's best model on failed lab experiments uses half the tokens of its predecessor for a clearly better score. We spend less and less to do more, but our agents gain autonomy the day their environment opens up. The question is therefore no longer which machine is the smartest, but how far we let the machine act alone. Three points converge on our side: Baidu's licence closes an option we thought was open, the ARTEX exposure shows a single initial approval is no longer enough, and Cuttlefish reminds us that mobile supervision is becoming a standard among hosting providers. Our immediate task is not to pick a more powerful model: it is to bound our agents by targets and quotas, and to keep the final say.