Operating systems · Compilers · Security · AI操作系统 · 编译器 · 安全 · AI

The layers your business runs on, engineered by the people who build them.

你的业务所依赖的底层,交给亲手打造它们的人。

For over a decade we have shipped production kernels, compilers and security systems, and now our own AI inference runtime. Customers bring us the work where a failure costs factories, funds or reputation.

十余年来,我们交付生产级的内核、编译器与安全系统,如今还有自研的 AI 推理运行时。客户把那些一旦出错就会让工厂停摆、资金受损、声誉崩塌的工作交给我们。

3yrs年

Validator keys protected without a single exposure

验证者密钥受保护,从未泄露

0incidents起事故

In more than two years of VED in production at our customers

VED 在客户生产环境运行两年多的安全事故数

471.1tok/s

SonicBoom code editing on a desktop RTX 3050

SonicBoom 在桌面级 RTX 3050 上的代码编辑速度

10+yrs年

Shipping kernels, compilers and security in production

在生产环境交付内核、编译器与安全系统

What we do我们做什么

Four disciplines, one accountable team.

四个领域,一个团队负责到底。

Hard problems cross layers. An AI deployment is also a security problem; a security fix is often a kernel or compiler problem. We work at all four, so the fix lands where the problem actually is.

难题总是跨层的。AI 部署同时是安全问题,安全修复又常常是内核或编译器问题。四层我们都做,所以问题在它真正所在的地方被解决。

01

Operating systems操作系统

Kernel, drivers and boot chain, down to the firmware.

内核、驱动与启动链,直到固件。

  • Kernel & driver engineering
  • Secure boot & firmware
  • Embedded Linux & IoT
  • Storage systems
  • 内核与驱动开发
  • 安全启动与固件
  • 嵌入式 Linux 与物联网
  • 存储系统
02

Compilers编译器

Toolchains and runtimes that generate and run your code.

生成并运行你代码的工具链与运行时。

  • LLVM, GCC & MLIR
  • Language runtimes & VMs
  • Cross-compilation
  • Toolchain security
  • LLVM、GCC 与 MLIR
  • 语言运行时与虚拟机
  • 交叉编译
  • 工具链安全
03

Security安全

Defense that holds against attacks nobody has seen yet.

面对尚未出现的攻击,依然守得住的防御。

  • VED kernel exploit defense
  • Hardening & code audit
  • Automotive & critical infrastructure
  • Privacy architecture
  • VED 内核漏洞利用防御
  • 系统加固与代码审计
  • 汽车与关键基础设施
  • 隐私架构
04 · AI

AI infrastructureAI 基础设施

Private AI on our own inference runtime and compiler.

基于自研推理运行时与编译器的私有化 AI。

  • Private LLM deployment
  • Inference performance
  • New hardware bring-up
  • AI security review
  • 私有化大模型部署
  • 推理性能工程
  • 新硬件适配
  • AI 安全审查
Products产品

Software we are known for.

让我们被人记住的软件。

Built by us, maintained by us, running in production. Open source, so you can read the code before you sign.

我们自己构建,自己维护,在生产环境中运行。全部开源,签约之前你可以先读代码。

Flagship · Linux kernel security旗舰产品 · Linux 内核安全

VED

A vaccine for the Linux kernel.Linux 内核的疫苗。

Most kernel defenses wait for a bug to be found and patched. VED stops the exploitation itself, so an attack on an unknown flaw fails the same way an attack on a known one does. It protects production fleets in manufacturing and finance today.

多数内核防护要等漏洞被发现、被修补。VED 直接阻断漏洞利用本身,攻击未知漏洞和攻击已知漏洞一样失败。它正在保护制造业与金融业的生产集群。

0-day kernel exploits内核 0day 利用Rootkits & persistenceRootkit 与持久化GPLv3
2+years without an incident on a smart factory's production nodes年零安全事故:智能工厂生产节点
2+years without an incident in an offline digital asset vault年零安全事故:离线数字资产金库
Security Lab ↗安全实验室 ↗
Automotive汽车

Breaking the exploit chain, link by link.

把攻击链一环一环打断。

A connected car is attacked through a chain: a radio, a chip's firmware, the infotainment kernel, then the vehicle network. Each link we harden is one an attacker can no longer use, from the SoC firmware up to the Linux kernel.

联网汽车遭受的攻击是一条链:无线模块、芯片固件、车机内核,再到整车网络。我们加固的每一环,攻击者都无法再利用,从 SoC 固件一直到 Linux 内核。

STANDARDS

AUTOSARAUTOSAR

In 2021 we submitted proposed changes to the security requirements of the AUTOSAR Adaptive Platform.

2021 年,我们向 AUTOSAR 提交了 Adaptive Platform 安全要求的修改建议。

SOC

Firmware first固件优先

Chipset firmware is opened up and reviewed, because the chain usually starts below the operating system.

芯片组固件要打开审查,因为攻击链往往从操作系统之下开始。

KERNEL

VED on boardVED 上车

Kernel privilege escalation on the head unit is a step attack contests put a high price on. VED is built to stop it.

车机内核提权是漏洞攻防赛事重金悬赏的环节,VED 正是为阻断这一步而生。

AI

AI that can ship能上车的 AI

Reproducible, Python-free inference with a small software bill of materials, ready for certification and audit.

可复现、无 Python 依赖、物料清单精简的推理,为认证与审计做好准备。

ISO/SAE 21434-aligned threat modeling, ECU hardening and secure in-vehicle communication. Detailed material is available to OEMs and Tier 1 suppliers on request.

符合 ISO/SAE 21434 的威胁建模、ECU 加固与车内通信安全。详细材料可应整车厂与 Tier 1 供应商要求提供。

Wi-Fi SoC exploit stageWi-Fi SoC 攻击阶段Mitigation缓解措施
SoC info leak: DMA to locate kernel symbolsSoC 信息泄露:通过 DMA 定位内核符号IOMMU / SMMU
Kernel hijack: overwriting the code劫持内核:覆写代码CFI
Heap shaping: crafting a predictable layout堆布局:构造可预测的内存布局Freelist randomization空闲链随机化
Kernel privilege escalation内核提权VED
Situational hardening: each stage of a long exploit chain is defeated on its own.情境化加固:长攻击链的每一个阶段都被单独击破。
Standards & public policy标准与公共政策
AUTOSAR · 2021Proposed changes to the security requirements of the Adaptive Platform.提交 Adaptive Platform 安全要求修改建议。
Hong Kong Monetary Authority · 2022香港金融管理局 · 2022Technical feedback on the e-HKD technical whitepaper, acknowledged in writing by the HKMA.就 e-HKD 技术白皮书提交技术意见,获金管局书面致谢。
Customer results客户成果

Measured in years without an incident.

用零事故的年数来衡量。

Customer names are withheld at their request.

应客户要求,客户名称不予公开。

Smart manufacturing智能制造Security安全

Kernel 0-days stopped on the factory floor

把内核 0day 挡在工厂车间之外

Connected production lines turned every device into a way in, and no signature can catch an unknown kernel exploit. VED now runs on every production node.

联网产线让每台设备都成了入口,特征库拦不住未知的内核漏洞利用。如今每个生产节点都运行着 VED。

2+ years年without a security incident零安全事故
Read the case →阅读案例 →
Digital asset custody数字资产托管Security安全

Rootkits stopped where no alert can get out

在发不出警报的地方拦住 rootkit

An offline vault cannot call for help, and its real threat is the implant that steals slowly. We built the threat model, deployed VED and tested it in production for two years.

离线金库无法向外求救,真正的威胁是悄悄窃取的植入物。我们建立威胁模型、部署 VED,并在生产环境测试两年。

2+ years年without a security incident零安全事故
Read the case →阅读案例 →
Ethereum staking以太坊质押Security安全

Validator keys safe for three years, on a lean security budget

三年密钥零泄露,安全开支保持精简

Six months of joint simulation on AWS proved the design. The operator then deployed it to production themselves, with our guidance.

在 AWS 上联合仿真六个月验证方案,随后客户在我们指导下自主完成生产部署。

3 years年keys never exposed密钥从未泄露
Read the case →阅读案例 →
Marketing · launching soon营销 · 即将发布AI

An AI marketing agent built on everything we make

一个集我们全部技术于一身的 AI 营销智能体

Daily market intelligence, content localized across cultures, product imagery and voice, generated on our own AI runtime and served by our own stack, with VED guarding every node.

每日市场情报、跨文化本地化内容、产品图与语音,由自研 AI 运行时生成、自有技术栈承载,VED 守护每一个节点。

Artanis · SonicBoom · velum
VED · AgentOSone system, every layer ours一套系统,每一层都是自己的
Read more →了解更多 →
HardenedVault AI LabHardenedVault AI 实验室

We don't wrap someone else's runtime. We wrote our own.

我们不套壳别人的运行时,我们自己写了一个。

SonicBoom is a native tensor runtime in C++ and CUDA with its own IR and kernels. No Python, one shared library, and output that stays bit-identical with speculative decoding on or off.

SonicBoom 是用 C++ 与 CUDA 编写的原生张量运行时,拥有自己的 IR 与算子。不依赖 Python,只有一个动态库,开关推测解码输出逐位一致。

Code editing, speculative decoding代码编辑,推测解码471.1 tok/s
vs llama.cpp's fastest speculative mode对比 llama.cpp 最快推测模式1.16×
Turns identical with speculation on and off推测开关输出一致的轮次10 / 10
Qwen2.5-7B-Instruct Q4_K_M · RTX 3050 8 GB · SonicBoom v0.0.6 · methodology
How we engage合作方式

From first call to production.

从第一次沟通到上线运行。

Fixed-scope assessments, project delivery or a long-term engineering retainer, under a license that fits your business.

固定范围的评估、项目交付或长期工程支持,以适合你业务的许可方式交付。

01

Assess评估

Threat model, architecture and performance review. You get a written report you can act on, with or without us.

威胁模型、架构与性能评审。你会拿到一份可执行的书面报告,用不用我们都能落地。

02

Build交付

We deliver the hardening, the deployment or the custom engineering, tested against your real workload.

我们交付加固、部署或定制工程,并用你的真实负载完成测试。

03

Operate运维

Ongoing support, updates and audits, by the engineers who wrote the code.

持续的支持、更新与审计,由写代码的工程师亲自负责。

Contact联系我们

Bring us the hard problem.

把最难的问题交给我们。

A private LLM that has to run on your own hardware, a fleet that has to survive the next kernel 0-day, a toolchain nobody else will touch. An engineer replies, not a sales desk.

要在自有硬件上运行的私有大模型、要扛住下一个内核 0day 的服务器集群、没人愿意碰的工具链。回复你的是工程师,不是销售。