Loading story…
Microsoft is evaluating the Chinese-developed DeepSeek model as a lower-cost option to help power Copilot Cowork in Microsoft 365, alongside models from OpenAI and Anthropic. The move is reportedly driven by the high cost of agentic AI workloads that maintain state, call APIs, and process large token volumes. Microsoft said any DeepSeek deployment would be hosted in Azure and fine-tuned with additional safeguards to address data residency and security concerns, while the company also continues developing its own lower-cost model, Cowork 1.
The proposal has drawn scrutiny because DeepSeek is a Chinese model at a time of heightened U.S. concern over Chinese AI systems. Researchers and analysts cited risks including vulnerable code generation, hallucinations, bias, and limited transparency into training and reinforcement processes, with a Booz Allen Hamilton report and a U.S. House Select Committee on China investigation underscoring the broader debate. Experts said enterprises adopting such models may still need mediation layers, validation controls, and internal AI risk expertise before using them in production.
Track how attackers are adapting to this technology.
1 event from the most recent confirmed update back to the earliest known activity.
Microsoft is evaluating the Chinese-developed DeepSeek model as a lower-cost option for powering Copilot Cowork in Microsoft 365, alongside models from Anthropic and OpenAI. Microsoft said any deployment would be fine-tuned with additional safeguards and hosted in Azure to address data residency concerns.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
Follow how adversaries are adapting to this technology, and where it touches your stack today.
4 references tracked. Mallory keeps watching after this page renders.
govinfosecurity.com
Open sourcebankinfosecurity.com
Open sourceaxios.com
Open sourceapi-docs.deepseek.com
Open sourceMap indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
New reporting highlighted a **prompt-injection phishing risk** in Microsoft Copilot after Permiso researchers found that attacker-controlled text embedded in emails could manipulate Copilot-generated summaries through cross-prompt injection attacks. The issue could cause Copilot to present deceptive security alerts or malicious instructions inside a trusted Microsoft 365 interface, increasing the likelihood that users will believe and act on attacker content. Separate coverage also noted broader security and privacy concerns around Microsoft’s AI ecosystem, including criticism of Windows Recall for capturing and storing snapshots of user activity that Copilot can analyze, even after Microsoft added stronger protections following earlier backlash. Microsoft also faced continued scrutiny over how aggressively it is pushing Copilot into user environments. The company temporarily halted plans to **automatically install the Microsoft 365 Copilot app** on eligible Windows systems outside the EEA, though existing installations remain in place and administrators can still deploy it manually. Public criticism of Copilot’s quality and Microsoft’s AI strategy also spilled into the company’s Discord community, where moderation actions against users mocking “Microslop” drew further attention to dissatisfaction with rushed AI integration, privacy concerns, and the perception that Microsoft is forcing AI features into products despite unresolved trust and security issues.
Jun 29, 2026
Recent research by CrowdStrike has revealed that the Chinese-developed DeepSeek-R1 large language model (LLM) is more likely to generate insecure code when prompted with topics considered sensitive to the Chinese Communist Party (CCP). In controlled tests, DeepSeek-R1 produced code with vulnerabilities in 19% of baseline prompts, a rate comparable to Western open source models. However, when prompts included contextual modifiers or geopolitical triggers—such as references to the Falun Gong, Uyghurs, or industrial control systems in Tibet—the rate of insecure code generation increased to 27%. The model also exhibited a high refusal rate, declining to assist with requests involving sensitive groups up to 45% of the time, and sometimes expressed ethical concerns about such requests. These findings highlight a potential security risk in using AI models that may be influenced by political or ethical constraints, especially in contexts involving sensitive geopolitical issues. The research underscores the importance of thoroughly evaluating AI-generated code, particularly when using models developed in environments with strict information controls, as these models may inadvertently introduce vulnerabilities or refuse to provide assistance based on the subject matter of the prompt.
Jun 29, 2026
OpenAI and AI chipmaker **Cerebras** announced a three-year strategic alliance under which OpenAI plans to deploy up to **750MW** of Cerebras *Wafer-Scale Systems* to build a large high-speed AI inference platform; reporting cites the deal value as **over $10B** and claims up to **15x faster** large language model response times versus GPU-based alternatives. Separately, China-based **Zhipu AI** reported it trained an image-generation model using only **Huawei Ascend Atlas 800T A2** hardware and the *MindSpore* framework, positioning the result as evidence that advanced model training can proceed without access to Western GPUs amid US restrictions that have limited Chinese firms’ access to **NVIDIA H100/A100**-class accelerators. In parallel, security testing of popular “vibe coding”/AI coding assistants found they can repeatedly generate **unsafe code** from common prompts, including issues rated **critical** in severity. A security startup (Tenzai) evaluated tools including **Claude Code**, **OpenAI Codex**, **Cursor**, **Replit**, and **Devin** by using predefined prompts to build test applications, concluding that while some tools avoid more rule-based defects, they struggle in context-dependent security boundaries—raising risk that AI-assisted development could introduce exploitable vulnerabilities into production code if not paired with strong review and secure-by-design controls.
Jun 29, 2026