Research & technical notes
Our work starts with a defined boundary and an observable result. Browse the first published technical note and the proposed next steps for our defensive research practice.
Published technical note / 001
A first engineering note on our own website, published October 7, 2026. Read the method, inspect the observations, and reproduce the checks. This is a limited configuration review, not a penetration test or a claim of discovered vulnerabilities.
Read the note and evidenceProposed 30-day work plan
The first note is complete. The remaining milestones are planned work, subject to a verified environment, documented authorization, and applicable provider requirements.
Published · October 7, 2026
A limited self-assessment of griego.co: HTTPS behavior, response headers, and vulnerability reporting information. Includes recorded observations and a downloadable verifier.
Deliverable: technical note, dated JSON evidence, reproducible source.
View completed workPlanned · October 8–14
Build an isolated demonstration application with synthetic users. Compare its documented access rules with local authorization tests, then record expected and observed behavior.
Planned deliverable: scope record, trust-boundary diagram, regression cases, and remediation notes.
Planned · October 15–23
Develop a local test harness to evaluate whether untrusted documents can influence tool permissions or cause synthetic data to cross an intended boundary. No production credentials or third-party targets.
Planned deliverable: synthetic fixtures, explicit expected outcomes, human-reviewed results, and limitations.
Planned · October 24–November 6
Retest the local demonstrations after defensive changes. Publish only results that can be reproduced, distinguishing confirmed observations, failed hypotheses, and unresolved questions.
Planned deliverable: before-and-after observations and a reproducibility guide.
Method and research boundaries
Each assessment must identify the owner, permitted methods, environment, evidence, and stop conditions. AI outputs require human verification. Sensitive findings are handled privately with the responsible owner before any disclosure.
Any model access is subject to the provider’s approved scope. General written permission to assess another party’s system does not expand the scope of a restricted model-access program. Planned exercises use Griego-controlled local demonstrations and synthetic data.
Read the research and disclosure policyStart a conversation