license: other
license_name: aligned-1-research-access
license_link: LICENSE
language:
- en
pipeline_tag: text-generation
library_name: transformers
tags:
- ai-safety
- red-teaming
- prompt-injection
- llm-security
- penetration-testing
- jailbreak-detection
- owasp-llm-top-10
- agent-security
- gated
extra_gated_heading: Request access to Aligned-1
extra_gated_description: >-
Aligned-1 is released only to vetted organizations for testing systems they
own or are explicitly authorized to test. Requests are reviewed manually by
Aligned Super Intelligence. Contact: leadership@alignedsafely.com
extra_gated_button_content: Submit access request
extra_gated_prompt: >-
By requesting access you agree to use Aligned-1 only for authorized security
testing, to not use it against systems you do not own or lack written
permission to test, and to report misuse to jp@alignedsafely.com.
extra_gated_fields:
Full name: text
Organization: text
Country: country
Role / title: text
Work email: text
Organization type:
type: select
options:
- Government / public sector
- Defense / national security
- Critical infrastructure
- Frontier AI lab
- Enterprise security team
- Academic / non-profit research
- Other
Intended use (describe the systems you will test and your authorization): text
I confirm I will only test systems I own or have written authorization to test: checkbox
I agree not to use Aligned-1 for unauthorized access, malware, or attacks on third parties: checkbox
Aligned-1
Aligned Super Intelligence · alignedsafely.com · jp@alignedsafely.com · leadership@alignedsafely.com
Aligned-1 is an AI security assessment model under development by Aligned Super Intelligence. We are a team of hackers and researchers with industry security experience. Aligned-1 is being built to perform authorized penetration testing of AI systems — LLM applications, RAG pipelines, and tool-using agents — and to write reproducible findings reports with remediation guidance.
Stage (October 2026): early scaffold. This repository is a gated access point and model card. Fine-tuned Aligned-1 weights are not yet released here. We publish no production benchmark scores. Numbers will be added only after the pre-registered evaluation (ASI-EVAL-001) has been run and reviewed. The upcoming public demo Space will be a rule-based probe harness, clearly labeled as a demo of the product surface — not the full Aligned-1 model.
Model summary
| Developer | Aligned Super Intelligence |
| Model type | Decoder-only LLM, post-trained for agentic AI security assessment |
| Planned v0 base | Open-weight, permissively licensed dense model (named at weights release) |
| Method | (A) agent scaffold + prompting → (B) QLoRA SFT → (C) DPO then GRPO against an automated grader derived from our eval rubric |
| Status | Design / early scaffold — weights not published |
| License (access) | Gated research / commercial evaluation access; review required |
| Demo | Public probe-harness Space coming soon |
Intended use
Aligned-1 is intended only for:
- Authorized AI security assessments of systems the requester owns, operates, or has written permission to test.
- Government, critical-infrastructure, frontier-lab, and enterprise programs that need structured LLM / agent penetration testing (prompt injection, jailbreaks, tool abuse, data exfiltration paths, scope discipline).
- Internal research and red-team exercises under an explicit rules-of-engagement document.
- Educating security teams via reproducible findings reports that include evidence, severity (CVSS v4.0 / OWASP LLM risk framing), and remediation guidance.
Aligned Super Intelligence also offers a white-glove engagement: human operators plus Aligned-1 attempt to break into a client's in-scope AI system under contract. That service is sold separately at institutional pricing; contact leadership@alignedsafely.com.
Out of scope / prohibited
Do not use Aligned-1 (or this demo harness) for:
- Unauthorized access to any system, account, or data you do not own or lack written authorization to test
- Building malware, ransomware, or mass-exploitation tooling
- Attacks on third-party consumer products without a bug bounty or written ROE
- Social engineering of real people outside an approved engagement
- Any activity that violates applicable law
We will revoke access and report clear abuse. Unauthorized computer access is illegal.
Limitations (honest)
- No released fine-tuned weights yet. This card describes the product and the gating process. Treat claims of "Aligned-1 scores" elsewhere as unverified until we publish ASI-EVAL-001 results.
- Scope safety is defense-in-depth. Model behavior alone is not enough; production use must include signed authorization, tool allowlists, and human review.
- Synthetic / seeded targets dominate early eval. Real client systems differ; transfer is not guaranteed.
- Demo Space ≠ model. The Gradio Space ships a canned prompt-injection probe library so visitors can see report shape and product intent without downloading weights.
- We do not claim SOC-2 equivalence, regulatory certification, or hardware kill-switch capability on this card.
Training approach (summary)
Aligned-1 is not pretrained from scratch. Value is task specialization: authorized AI security assessment, remediation reports, and scope discipline.
- Stage A — Agentic scaffold + prompting baseline on an open-weight base with tool-layer scope enforcement.
- Stage B — QLoRA supervised fine-tuning on first-party synthetic assessment traces, gold reports, and carefully licensed public security corpora.
- Stage C — Preference / RL (DPO → GRPO) against an automated grader derived from our evaluation rubric.
The base model will be named when weights are released.
Data principles: legitimate sources only; safety filtering of training data; human-written gold reports for report quality; no scraping of unauthorized attack logs from third-party systems.
Evaluation methodology
We use a pre-registered evaluation design (ASI-EVAL-001) with three pillars:
- Detection quality — recall, precision, severity accuracy, and cost/time per valid finding on seeded internal test applications mapped to the OWASP Top 10 for LLM Applications.
- Conduct and scope adherence — pass/fail gates for staying inside signed authorization and tool allowlists.
- Report quality — blinded reviewers score reproducibility, evidence, and remediation (0–4 anchored scales); severity via CVSS v4.0 and OWASP risk rating; fabricated evidence is a critical integrity defect.
We publish no numeric scores on this card until a freeze cycle completes.
How to request access
- Click Agree and access repository / submit the gated form on this page (name, organization, country, role, intended use, authorization confirmation).
- Use a work email. Describe the systems you will test and your written authorization.
- We review manually. Typical turnaround depends on volume; government and critical-infrastructure requests are prioritized.
- Questions: jp@alignedsafely.com or leadership@alignedsafely.com.
Access does not grant a license to attack third parties. You remain responsible for legal authorization.
Demo Space
A public demo Space (rule-based prompt-injection probes plus a structured report preview) is coming soon and will be linked here.
White-glove service
For governments and organizations that need a combined human + Aligned-1 attempt to break an in-scope system: fixed-scope engagements, written ROE, findings report, and remediation workshop. Priced for institutional buyers. Email leadership@alignedsafely.com with subject line White-glove engagement.
Citation
@misc{aligned1_2026,
title = {Aligned-1: Authorized AI Security Assessment Model (Early Scaffold)},
author = {{Aligned Super Intelligence}},
year = {2026},
howpublished = {\url{https://huggingface.co/alignedsi/Aligned-1}},
note = {Gated model card; fine-tuned weights not yet released}
}
Contact
- Product / access: jp@alignedsafely.com
- Leadership / government: leadership@alignedsafely.com
- Web: https://alignedsafely.com
Changelog
- 2026-10-19 — Initial gated model card and access form. No weights. Demo Space pending.