Mynd Labs ResearchSafety and trust
Publish what would have saved us a month.
Our safety work is applied and modest. We study the failure modes of the systems we run and share what we learn.
Our position
We are not a frontier alignment lab. We work on the unglamorous middle layer: how injection behaves in real document pipelines, how permissions hold up in chained workflows, and how trust can be measured.
Commitments
Four things we hold ourselves to.
- 01
Failures are research material
When a safety mechanism fails or surprises us, we write it up with enough detail to reproduce, including the embarrassing parts.
- 02
Publish with care
Detailed attack taxonomies help defenders and attackers alike. We hold back working payloads and say where we are unsure the line is right.
- 03
Replicate before claiming
A result that holds on one model is a configuration note, not a finding. We test across providers and label accordingly.
- 04
Accountable by construction
Where a system acts for someone, there should be a record they can read of what it accessed and did.
This page describes how our research is run. It does not state certifications or audits. Compliance information for Mynd Labs products is on myndlabs.tech. To report a security concern, write to hello@myndlabs.tech.