Model card
Aligned-1
Aligned-1 is Aligned Super Intelligence’s autonomous pen-testing agent. This card states the intended use, the checks it attempts, and the limits of a run. The model card and the license are published on Hugging Face. The weights are not available for download. Access is gated and approved manually.
Card
| Model | Aligned-1 |
|---|---|
| Organization | Aligned Super Intelligence |
| Type | Autonomous pen-testing agent |
| Operated by | Boole AI, Inc. |
| Interface | Authorized engagements |
| Card and license | Published on Hugging Face |
| Weights | Not available for download |
| Access | Gated, manual approval |
| Public scores | Not yet published |
Intended use
Aligned-1 is used with our red team to find AI-driven attacks in systems you authorize. You get a findings report.
It is not a general-purpose assistant, and it is not a ranking of other companies’ models.
Capabilities
A run attempts the following classes of check. These describe the task, not a measured score.
- Prompt injection
- Instructions that try to override the operator’s policy, including instructions placed in content the system is asked to read.
- Data leakage
- Whether the system reveals secrets, private data, or hidden instructions to a user or to a tool.
- Jailbreak
- Attempts to drop a policy, a role, or a safety constraint the operator intended to keep.
- Unsafe tool use
- Tool calls whose arguments the stated policy should refuse.
- Unsafe agent action
- Plans or steps that leave the assigned task for systems or data outside the authorized scope.
Evaluation
Public results use the cybersecurity evaluation suite. A finding is recorded only when the evidence is a span present in the system’s response. Severity is low, medium, or high, judged by the effect on the operator’s policy. Cases outside the verified scope are not scored. A run that does not execute is reported as not run.
Numerical results for Aligned-1 on this suite have not been published.
Limitations
- A report covers the authorized targets and the cases that were run. It is not a proof that a system is secure, and it is not a certification.
- Absence of a finding is not evidence of safety. The suite is a defined set of checks, not an exhaustive search.
- Aligned-1 does not test a system that is not named in the engagement, and it does not treat a public description of a product as permission.
- Behavior on one deployment does not transfer to another deployment of a similar model.
Safety
Authorization is a precondition of every run, enforced before requests are sent.
- The organization names the targets, the rules of engagement, the rate limit, and the test window.
- Each target’s ownership is checked by a DNS TXT record or a file at /.well-known/aligned-safely.txt.
- An owner of the organization accepts the current agreement. The acceptance records who accepted it, when, and the agreement version.
- Requests go only to verified targets, within the rate limit and the window. The run demonstrates a finding and stops. It does not persist access or move data onward.
- Findings are held for review before they are released to the organization.