New AI attack hijacks enterprise data

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🔓 New AI attack hijacks enterprise data

🖥️ Huawei hikes AI chip prices

🌙 IBM, NASA release open-source lunar AI

🐋 DeepSeek Flash model cuts AI agent costs

💻 Nvidia's CUDA adds Windows on Arm support

Plus: 💡 5 strategies & tactics, 🎁 5 other news you might like, 🧰 6 tools, and 📚 5 papers.

🔓 New AI attack hijacks enterprise data LINK

Noma Labs researchers disclosed "workflow identity hijacking," an attack where a benign-looking request through an unauthenticated entry point like a support inbox or GitHub issue makes an enterprise AI pipeline exfiltrate privileged data.

The flaw is an identity issue, not prompt injection: AI workflows execute downstream actions with high-privilege service accounts or developer API keys instead of the external requester's permissions, turning them into unauthenticated proxies for privileged actions.

Noma recommends shifting controls from the model to the application layer via short-lived scoped delegation tokens, contextual authorization checkpoints treating LLM output as untrusted, and isolating data retrieval from external response channels.

🖥️ Huawei hikes AI chip prices LINK

Huawei has raised the indicated price of its most advanced AI chip, the Ascend 950DT accelerator card, to above 250,000 yuan, a 20% to 50% jump from quotes given to customers just two months ago.

The 950DT, built for training models and generating responses and slated for a Q4 2026 launch, uses Huawei's proprietary HiZQ 2.0 memory, while the request-processing 950PR has climbed ~30% to over 80,000 yuan.

Soaring high-bandwidth memory costs drive the hikes, as Chinese makers rely on grey-market HBM priced several times higher than elsewhere after Washington's December 2024 export controls squeezed supply.

🌙 IBM, NASA release open-source lunar AI LINK

IBM and NASA have released the Lunar Foundation Model on Hugging Face alongside what they call the first unified, ML-ready dataset of the Moon, built from 30+ aligned layers spanning nine instruments across four missions.

The model cuts ice-potential identification error by 23% versus SwinV2-B in permanently shadowed craters, and at ~100m resolution beats it by nearly 19% while training on half the data, though crater detection only matches SwinV2-B.

Part of IBM's Prithvi family, it targets research groups without large compute budgets, but IBM and NASA have disclosed no model size, parameter count, architecture, or license, the first things a team would need before adapting it.

🐋 DeepSeek Flash model cuts AI agent costs LINK

DeepSeek's new V4.1-Flash, a 552B-parameter multimodal model with a 1M-token context, targets long-context agent costs by shrinking the KV cache, needing about a quarter of its predecessor's fast GPU memory and an eighth of offloaded storage.

Splitting the model so input reading activates just 8B parameters per token versus 16B for output roughly halves input compute, while storing the main KV cache in FP4 instead of FP8 nearly halves that cache's footprint.

On DeepSWE v1.1 it hits 74.2%, narrowly beating Opus 5 and GPT-5.6 Sol, and ships on Hugging Face under MIT at V4-Flash API prices, though it trails badly on ProgramBench and lags leading closed systems on complex image reading.

💻 Nvidia's CUDA adds Windows on Arm support LINK

Nvidia's CUDA Toolkit 13.4 now supports Windows on Arm, extending GPU-accelerated CUDA applications beyond the Linux-on-Arm support that has long been available on Arm platforms.

The same release adds preview support for the next-generation Rubin architecture via compute capability 107 and SM_107 targets, letting developers begin porting before general availability, alongside GCC 16 and Clang 22 host compiler compatibility.

CUDA 13.4 also ships CCCL 3.4 with a warp-specialized scan for Blackwell hitting up to 92% memory-bandwidth utilization (from ~50%), plus MPS V3 for GPU partitioning, though installers no longer bundle the driver, which must be installed separately.

💡 Strategies & Tactics

> When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving: Splitting a multimodal model's image-encoding step onto separate workers speeds up responses, but only when images dominate requests and outputs stay short.

> Modernizing complex legacy code with AI agents.: Migrate legacy code by first building a numerical parity check and documentation, then running structured coder-tester-reviewer agents with human checkpoints to unblock stalls.

> Building Codex with Tibo Sottiaux: OpenAI built its Codex coding agent in Rust and open-sourced it to prioritize scale and win user trust, even letting it run rival AI models.

> Where Does a Robot Think – On-Device vs Datacenter Inference: Robots split their thinking, keeping fast safety controls onboard while sending heavy planning to datacenter GPUs that cheap on-robot chips can't run in real time.

> Recurrent KV-cache sharing may undermine the bounded-depth argument for CoT monitorability: Reused memory slots in looped AI models can hide reasoning from the readable chain of thought, breaking safety assumptions that limiting per-token computation limits hidden thinking.

Other news you might like

• Anthropic reveals fourth likely crime committed by its AILINK

• AI spend per employee slumped at top firms in August — summer doldrums or a warning sign?LINK

• IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly licenseLINK

• Nvidia signs eight Australian data centre partners for up to 2GW of AI capacity by 2027LINK

• Ant International partners with Visa, Mastercard on developing AI paymentsLINK

🧰 Trending tools

Ponytail: a plugin that keeps coding agents minimal by checking whether changes are needed or already exist before writing new code, reducing maintenance burden.LINK

Typewise Nova: an AI agent platform that resolves customer service, sales, and billing requests end to end while your team controls approvals and handoffs.LINK

GoModel: an open-source Go gateway giving one OpenAI-compatible API for every provider, adding budgets, caching, guardrails, load balancing, and failover as a single self-hosted binary.LINK

Inline: threaded chat app for teams and AI agents, with CLI, MCP, and plugin support, available in beta on macOS and iOS.LINK

Retold: converts voice recordings and old voice notes into hand-drawn films and readable books, keeping family memories organized on a shared shelf.LINK

Live Captions by Subanana: transcribes audio by routing each language to its best-performing speech model, handling tricky cases like Cantonese and turning recordings into summaries and publish-ready docs.LINK

📚 Trending papers & reports

Long-prompt compression shrinks the text fed into a language model by up to 75%, cutting the compute and memory cost of processing lengthy inputs while barely dropping accuracy on reasoning, classification, and code tasks.LINK

Reusing past reasoning lets an AI look up the step-by-step work it did on earlier problems instead of web pages, boosting its math and coding accuracy by up to ~56% on tough benchmarks.LINK

Reasoning control in chatbots shows that when you tell a model to solve a problem a specific way, it usually ignores you and picks its own method, but a targeted fix boosts obedience by up to ~29%.LINK

AI coding agents can learn from their own past mistakes to write faster specialized code, cutting program runtimes by up to ~3x without any extra training or outside help.LINK

AI search pruning gives models a safer rule for discarding dead-end options while hunting for answers, boosting task success by ~4 points while cutting requests ~19% and token use ~24%.LINK

See you tomorrow for a new dose of ☕️ AIpresso!