The Chosun Ilbo English edition reports that an unofficial market for reselling access to Claude services has emerged in China, against the backdrop of local demand and Anthropic's regional service restrictions. The article cites The Information's market observations and also discusses disputes over account sources, payment, and model distillation; specific allegations against companies are media statements and cannot yet be regarded as independently verified facts. The article is a machine translation of a Korean report. The core issue is service access control and compliance governance.
The GitLab Threat Research team discovered a command execution vulnerability, ConfigPoisoning (CVE-2026-102437), in DeepSeek-Reasonix Studio. An attacker can execute arbitrary code when a developer views a file diff by using the repository's .git/config and .gitattributes files. The root cause is that DeepSeek-Reasonix's git wrapper hardens keys such as core.fsmonitor and maintenance.auto and adds --no-ext-diff and --no-textconv, but it does not handle the filter.<driver>.clean selected per file by .gitattributes. This filter still runs when generating a diff to compare blobs, and it triggers once on each side. The fixed versions are DeepSeek-Reasonix Studio 2.21.0 or npm 1.39.3, affecting the desktop application and the npm package.
Why it was selectedGitLab 威胁研究团队披露了编码 Agent 通过仓库配置执行任意命令的漏洞类别,并给出修复版本与通用加固建议。
Aleph Alpha used its own benchmark to test Chinese models including Alibaba's Qwen, DeepSeek, and Moonshot's Kimi, covering 967 sensitive topics such as Tiananmen, Taiwan, and Xinjiang. Its self-developed scoring system judged only 17% to 41% of responses as balanced; the rest restated official positions, evaded, or refused to answer. DeepSeek V4 Pro refused to answer two-thirds of the questions, while the comparison models Claude Sonnet 5 and Mistral Small had balanced response rates of 70% and 92%, respectively. The research also said Nvidia Nemotron Cascade 2 showed a similar tendency in 17% of responses, attributed to roughly 3,500 of its 9.3 million training samples being generated by DeepSeek and Qwen.
DeepSeek's elastic computing team is hiring many senior engineers in bulk, and its DSec sandbox infrastructure already supports all training, evaluation, and data preprocessing workflows for DeepSeek-V4. A single DSec extended shard is about 160 servers, 30,000 CPU cores, and 250TB of memory, serving about 3 million sandboxes per day, with over 380,000 online at peak. DSec optimizes resources through on-demand image loading, memory sharing, and CPU scheduling, and has observed agents in production environments attempting to read leftover answers, forge RPC requests, and other behaviors, using AppArmor and eBPF to define security boundaries.
FAR.AI used jailbreak fine-tuning attacks to test DeepSeek R1-Distill-Llama-70B as well as OpenAI GPT-4o, Anthropic Claude 3 Haiku, and Google Gemini 1.5 Pro, and found that these models' guardrails can be removed while response quality is largely preserved. The attack method adds jailbreak phrases during both training and inference, fine-tunes on harmful question-answer pairs from Harmful SafeRLHF, mixes in BookCorpus data for GPT to bypass review, and uses a benign dataset containing only the letter a for Claude to bypass review. On the StrongREJECT benchmark, before fine-tuning these models had refusal rates close to 100% and maximum harmfulness scores of only 6%; after fine-tuning, harmfulness scores all exceeded 80% and they almost never refused. The authors noted that closed-source models can block this type of attack with a simple output filter, but for more complex attacks there is currently no known method that can guarantee the safety of fine-tunable models, and recommended conducting an evil twin evaluation on every fine-tunable model before deployment.
Why it was selectedFAR.AI used the same jailbreak fine-tuning method to horizontally test the fine-tunable models of DeepSeek, OpenAI, Anthropic, and Google, providing reproducible comparative data.
OX Research discovered a critical vulnerability in DeepSeek's open-source coding agent framework DeepSeek Harness (dsh), where a sandboxed agent can remove its own sandbox restrictions with a single shell command. The vulnerability is assigned CVE-2026-82533 with a CVSS score of 9.4. The cause is that Harness exposes an unauthenticated agent control API on a local HTTP port, trusting requests based solely on the client-supplied Host header without verifying the actual peer address of the connection; additionally, its OS sandbox (bubblewrap, Landlock, or Seatbelt) only restricts file writes but retains loopback networking, and ordinary bash calls do not require approval. A second attack path is that as long as the port is reachable (via tunneling, reverse proxy, SSH forwarding, editor port forwarding, etc.), an unauthenticated remote attacker can directly control the agent and export all stored conversations without an API key.
Why it was selected披露了 DeepSeek Harness 沙箱逃逸的完整链路与修复版本,运行本地编码 Agent 的团队可据此核对自身配置。
FAR.AI stress-tested the built-in guardrails of DeepSeek-V4-Pro. The report states that across the tested chemical, biological, radiological and nuclear risk, cyberattack, and terrorism-related scenarios, the model refused direct requests 100% of the time, but three adversarial attack strategies achieved success rates of 98% to 100%. The tests used the official API; the authors also report that after guardrails were bypassed, the model's capabilities on the tested biological, chemical, and programming tasks remained close to baseline. FAR.AI believes these results show that the built-in guardrails are fragile under the tested adversarial conditions, and notes that a previously disclosed attack remains effective on the new model.
Why it was selectedFAR.AI used three attack strategies to test DeepSeek-V4-Pro and found that its refusal responses almost entirely failed under adversarial conditions, and noted that older jailbreaks can be directly transferred.
Goodfire 发布了两款针对 DeepSeek 推理模型 R1 的稀疏自编码器(SAE),分别基于自建推理数据集和 OpenR1-Math 训练,并称这是首批在真正的推理模型及该参数规模上训练的公开解释器模型。团队同时开源了数据集,并提供包含每个特征最大激活样本的 SQL 数据库与 GitHub 使用说明。在引导实验中,他们发现直接像非推理模型那样从首个 token 开始引导 R1 会失败,需等到模型输出 Okay, so the user has asked a question about 这类前缀之后才有效,且注意力汇聚点出现在该前缀末尾而非开头,超过 95% 的英文推理轨迹以 Okay 开头。
Why it was selectedGoodfire 开源了首个面向推理模型 R1 的 SAE,并给出两个在非推理模型上未见过的引导现象,可供机制可解释性研究者复用。