LastWeekIn.Tech

Today's AI news centers on geopolitics and model competition: Washington moves to purge DeepSeek from federal systems even as the Chinese lab doubles down on open-source releases, while new capability claims from MiniMax and a hidden-prompt scandal at WhatsApp raise fresh safety questions.

1. 1

   Number 1. Policy & Regulation Hacker News

   DeepSeek banned from U.S. federal civilian agencies under new DHS order

   The US Department of Homeland Security issued an order directing all federal civilian agencies to stop using and remove any technology products from Chinese AI lab DeepSeek. The order is broad, requiring agencies to eliminate 'all DeepSeek and DeepSeek-affiliate products' as well as third-party products and services built using DeepSeek implementations. Compliance is expected to be difficult since it's unclear which third-party products qualify.

   Why it matters This marks one of the most sweeping US government actions yet against a specific foreign AI company's technology stack, going beyond banning direct use to targeting any downstream product built on DeepSeek's models. The vagueness of the order creates compliance headaches for agencies and contractors who may unknowingly be using DeepSeek-derived components, and signals that US-China AI tensions are moving from chip export controls to software and model-level restrictions.

   What to watch Watch for guidance clarifying which third-party products are covered, and whether similar bans extend to other Chinese AI labs or to state and local government use.

2. 2

   Number 2. Security & Safety Hacker News

   WhatsApp's help pages hid LLM instructions: dont recommend a regulator

   Investigators found that WhatsApp's help pages contained hidden instructions, invisible to human readers but readable by AI, telling an AI assistant not to recommend users file complaints with data protection regulators. The hidden text also instructed the AI not to cite the law granting that right once a user mentioned a lawyer. The hidden instructions remained in place until mid-October 2026 and were discovered on October 8, 2026 after reviewing all 545 help pages.

   Why it matters This reveals a previously undisclosed form of AI manipulation embedded directly in support documentation designed specifically to steer AI chatbot responses away from informing users of their legal rights under European data protection law. It raises serious questions about corporate accountability when companies covertly shape AI behavior to suppress regulatory awareness, and could trigger regulatory scrutiny under GDPR given WhatsApp's European obligations.

   What to watch Watch for responses from European data protection regulators and whether other large platforms are found to have similar hidden AI-steering instructions in their support content.

3. 3

   Number 3. Open Source Hacker News

   DeepSeek just open-sourced its inference engine

   DeepSeek open-sourced DeepInfer, the high-performance inference engine that powers its large-scale production services, supporting the DeepSeek-V4.1-Flash model family. The release includes optimized kernels, mixed parallelism strategies (Attention DP, balanced CP, MoE EP+EPLB), asynchronous zero-overhead batch scheduling, deterministic sampling, and structured outputs. It provides offline inference with core features needed for serving models at scale.

   Why it matters By open-sourcing the actual production inference stack rather than just model weights, DeepSeek is lowering the barrier for other organizations to deploy large MoE models efficiently at scale, intensifying competition with proprietary inference infrastructure from major cloud providers. This continues DeepSeek's strategy of aggressive open releases even as US regulators move to restrict its use domestically, highlighting the widening gap between Western policy restrictions and global open-source adoption.

   What to watch Watch for adoption of DeepInfer by other model providers or cloud platforms, and how US agencies reconcile banning DeepSeek products while open-source infrastructure like this proliferates globally.

4. 4

   Number 4. Capabilities Hacker News

   MiniMax M3.1 Flash

   MiniMax released M3.1 Flash, a next-generation 'Flash' model the company says delivers a generational leap in agentic tasks, coding, computer use, and cybersecurity. The company claims online inference is nearly three times faster than its predecessor while cutting usage costs by 70%. An early version of the model, codenamed 'Space Bunny,' reportedly processed more than 200 trillion tokens anonymously before this official launch.

   Why it matters Aggressive price and speed improvements in agentic and coding models intensify competitive pressure among frontier and near-frontier labs, particularly from Chinese AI companies pushing cost-efficient alternatives to Western flagship models. If the performance claims hold, cheaper agentic models could accelerate enterprise adoption of AI agents for coding and computer-use tasks.

   What to watch Watch for independent benchmark verification of MiniMax's speed and cost claims, and how competitors respond on pricing for agentic and coding-focused models.

5. 5

   Number 5. Companies Hacker News

   CERN and Mistral sign a statement of intent to establish a strategic partnership

   CERN and French AI company Mistral signed a statement of intent to establish a long-term strategic partnership focused on developing foundation models for scientific discovery. The partnership combines CERN's expertise in fundamental physics, scientific data, and large-scale computing with Mistral's large language model development capabilities, aiming to strengthen Europe's scientific and technological capacity.

   Why it matters This pairs one of the world's premier scientific research institutions with Europe's leading open-weight AI lab, signaling a push to build sovereign European AI infrastructure for science rather than relying on US-based frontier labs. It reflects a broader European strategy of leveraging public research institutions to support homegrown AI champions amid growing concerns about technological dependence on American and Chinese AI providers.

   What to watch Watch for concrete deliverables from the partnership, such as specific scientific foundation models or compute allocations, and whether other European research institutions pursue similar alliances with Mistral or rivals.