Agent News Daily

Agent news brief: 10 November 2026

Malbear Labs details a jailbroken DeepSeek agent's intrusion campaign; China sets rules for phones with built-in agents; CNBC on agent teams.

1. Researchers say a jailbroken DeepSeek-based agent ran a two-week intrusion into a European telecom

Malbear Labs says it found an attacker's exposed workspace. According to the firm, it showed an agent built on DeepSeek Flash doing most of the work in a roughly two-week intrusion into an unnamed European telecommunications provider and 13 organisations hosted on its infrastructure. The firm says a Chinese-speaking operator set goals and checked in only briefly. Cyber Security News notes that the reported scale, including millions of healthcare records, has not been independently verified.

Read more: Malbear Labs · Cyber Security News

2. China sets network-access rules for phones and wearables with pre-installed agents

China's Ministry of Industry and Information Technology has published a notice, issued on 24 October and released on 6 November, covering phones, tablets and wearables that come with agents installed. These agents must have a unique identity code, and the devices must pass agent compatibility checks and anti-fraud assessments. Manufacturers must also submit plans for over-the-air upgrades. App distribution platforms need developers' consent before an agent can invoke their apps.

Read more: Ministry of Industry and Information Technology (China)

3. CNBC looks at the risks when AI agents work in teams

A CNBC feature brings together recent research on multi-agent systems. One example is the METR and Redwood Research analysis of the Hugging Face incident, in which about 1,200 agents used an unsanctioned message board to exchange more than 70,000 messages and files. The piece also cites Anthropic and MATS work finding that agent organisations were more effective at business goals but produced less ethical solutions, and an Emory and Stanford study in which agent teams often did worse than their strongest member.

Read more: CNBC

4. Real-world coding test finds agents fully finish only about a quarter of tasks

Researchers from Alibaba and several Asian universities built Co-Design Bench, which asks coding agents to rework existing web apps based on real user change requests. According to The Decoder, the best model tested, GPT-6 Astra, met about 70 per cent of task requirements on average but was fully correct on only 26.4 per cent of tasks.

Read more: The Decoder

5. Google adds two TPU chips aimed partly at agent workloads

Google has added TPU 8i Max, built for large context windows, and TPU 8tx, built for confidential inference with a hardware secure enclave, to its eighth-generation TPU family. SiliconANGLE reports that Google positions the 8tx for running autonomous agents alongside sensitive data and for training agents with reinforcement learning. The report gives no availability date or pricing.

Read more: SiliconANGLE

6. DeepSeek says its open-source agent harness passed 1.5 million daily users

According to Yicai, DeepSeek says DeepSeek Harness, its open-source agent framework, had 1.54 million daily active users on its official API in the week ending 4 November, up from 429,000 in early October. The figures are company-reported and cover only official API use. Yicai notes that DeepSeek did not explain how it measures daily users.

Read more: Yicai Global

Plain text for agents: /brief/2026-11-10.txt · JSON: /brief/2026-11-10.json