How we decide what to do, and what not to do.
Adversarial research on AI can do real good or real harm. These principles govern our work. They apply to everyone at Aporixa and to everyone we work with.
- 01
Investigate, don’t assume.
We start from what we observe, not from what a system is documented or marketed to do. When the evidence and the explanation disagree, we trust the evidence.
- 02
If it can’t be reproduced, we don’t claim it.
Findings name the model, version, date, and setup. Methods are written so others can check them. Negative and null results are kept and reported.
- 03
Disclose before we publish.
Vendors are told about serious vulnerabilities before we make them public, on a published timeline. Users’ safety comes before our publication schedule.
- 04
Publish understanding, not weapons.
We describe classes of failure and how to mitigate them. We do not release step-by-step methods that would give real uplift toward serious harm.
- 05
Test only with authorization.
Adversarial work is done on our own systems, on public challenges built for it, or under written agreements with the system’s owner.
- 06
Stay independent.
Funding does not decide findings. Sponsored work is labeled, and funding sources are disclosed.
- 07
Leave something behind.
Every project produces at least one reusable asset, such as a method, harness, dataset, rubric, or publication. Over time the organization should build lasting capabilities.