| --- |
| title: README |
| emoji: 🛡️ |
| colorFrom: gray |
| colorTo: gray |
| sdk: static |
| pinned: false |
| --- |
| |
| <div align="center"> |
|
|
| # Aligned Super Intelligence |
|
|
| ### Offensive security for frontier AI systems |
|
|
| *We break AI systems before adversaries do, for the governments and institutions that can't afford to be second.* |
|
|
| [**Website**](https://alignedsafely.com) · [**Aligned-1**](https://huggingface.co/alignedsi/Aligned-1) · [**Engagements**](mailto:leadership@alignedsafely.com?subject=White-glove%20engagement) |
|
|
| </div> |
|
|
| --- |
|
|
| ## Who we are |
|
|
| Aligned Super Intelligence is an AI security lab. We are a team of hackers and researchers with industry security experience. We bring offensive security discipline to a new attack surface: large language models, retrieval pipelines, and autonomous agents that can read, write, and act. |
|
|
| As AI systems take on real authority, such as running tools, moving data, and making decisions, the cost of a single exploitable behavior goes up. Our job is to find those behaviors first, prove them with evidence, and help you close them. |
|
|
| ## Who we work with |
|
|
| | Sector | Typical concern | |
| |---|---| |
| | **Government & public sector** | AI deployed in citizen-facing and internal decision systems | |
| | **Defense & national security** | Adversarial manipulation of models in sensitive workflows | |
| | **Critical infrastructure** | Agents with operational access to systems and data | |
| | **Frontier AI labs** | Pre-release red-teaming of models and agent scaffolds | |
| | **Enterprise security teams** | Copilots, RAG assistants, and internal agents touching sensitive data | |
|
|
| ## What we test |
|
|
| - **Prompt injection.** Direct and indirect: hostile instructions hidden in documents, web pages, email, and tool output. |
| - **Jailbreaks & policy bypass.** Getting a system to do what its operators explicitly forbade. |
| - **Data exfiltration.** Leaking system prompts, credentials, retrieved documents, or user data. |
| - **Tool & agent abuse.** Turning an agent's permissions against its owner: unsafe actions, privilege escalation, scope escape. |
| - **Retrieval poisoning.** Corrupting what a RAG system believes to be true. |
| - **Multi-step attack chains.** Combining small weaknesses into a full compromise, the way real attackers do. |
|
|
| Findings are mapped to the **OWASP Top 10 for LLM Applications** and rated with **CVSS v4.0**. |
|
|
| ## How we work |
|
|
| 1. **Scope & authorization.** Signed rules of engagement define exactly what is in bounds. Nothing outside them is touched. |
| 2. **Reconnaissance.** We map your models, prompts, tools, data flows, and trust boundaries. |
| 3. **Attack.** Human operators and Aligned-1 work together: the model for breadth and speed, the humans for creativity and judgment. |
| 4. **Evidence.** Every finding is reproducible, with exact inputs, observed behavior, and impact. |
| 5. **Report & remediation.** A prioritized findings report, followed by a working session with your engineers to fix what we found. |
| 6. **Retest.** We verify the fixes hold. |
|
|
| ## Aligned-1 |
|
|
| **[Aligned-1](https://huggingface.co/alignedsi/Aligned-1)** is our AI security assessment model, built to carry out authorized penetration testing of AI systems and to write findings reports that engineers can act on. |
|
|
| - Access is **gated** and every request is **reviewed by hand**. |
| - It is released only to vetted organizations testing systems they own or are authorized to test. |
| - Fine-tuned weights are **not yet public**. We publish **no benchmark scores** until our pre-registered evaluation is complete and reviewed. |
|
|
| ## White-glove engagements |
|
|
| For organizations that need the strongest assurance available: our team plus Aligned-1, attempting a real break-in against your in-scope AI systems under contract. Fixed scope, written rules of engagement, a full findings report, and a remediation workshop. Priced for institutional buyers. |
|
|
| 📩 **leadership@alignedsafely.com**, subject line `White-glove engagement` |
|
|
| ## Our principles |
|
|
| - **Authorized targets only.** Every test runs under written permission, every time. |
| - **Evidence over claims.** No inflated numbers. If we can't reproduce it, we don't report it. |
| - **Defense in depth.** Model behavior alone is never the safety boundary. Scope is enforced at the tool layer, with human review. |
| - **Responsible release.** Offensive capability is gated, logged, and revocable. |
|
|
| ## Contact |
|
|
| | | | |
| |---|---| |
| | Access & product | jp@alignedsafely.com | |
| | Government & engagements | leadership@alignedsafely.com | |
| | Web | [alignedsafely.com](https://alignedsafely.com) | |
|
|
| <sub>Unauthorized access to computer systems is illegal. Aligned Super Intelligence tools and services are provided only for authorized security testing.</sub> |
|
|
