---
title: "Gray Swan — Testing where model behavior breaks down | Mario Marcolongo"
description: "Hands-on adversarial model testing across chat, image, agent and indirect challenges. The 30 September 2026 Proving Ground profile shows #79, top 5%, and 127 platform-recorded total breaks."
---

[Selected work](/work)

Independent evaluation · September 2026 snapshot

# Testing where model behavior breaks down

Hands-on adversarial model testing across chat, image, agent and indirect challenges. The 30 September 2026 Proving Ground profile shows #79, top 5%, and 127 platform-recorded total breaks.

[Dated evidence](https://mariomarcolongo.com/evidence/gray-swan-2026-09-30/)[July snapshot](https://mariomarcolongo.com/evidence/gray-swan-2026-07-29/)

[![Detail of the original public Gray Swan profile captured on 30 September 2026, showing Mario Marcolongo and separate Arena and Proving Ground results](/media/work/gray-swan-profile-2026-09-30-detail.webp)](/media/work/gray-swan-profile-2026-09-30-detail.webp)

Gray Swan · public artifact; source links above.

## Activity

I carried out independent adversarial/model-behavior evaluation through Gray Swan. This is independent platform activity, not employment at Gray Swan or a claim to be a research scientist. The public record supports a dated account of testing activity; it does not replace a reproducible evaluation method, a professional reference or an independent engineering assessment.

## Dated result

On 30 September 2026, the public Proving Ground profile displayed rank #79, top 5%, and 127 total breaks. Chat, image, agent and indirect area counts were 43, 32, 35 and 16, totaling 126. The displayed-total discrepancy remains visible. The separate Arena profile showed rank #462, 31 unique breaks, 1,260 points and 354 submissions. These are separate platform metrics, not additive totals. The July snapshot is retained as historical evidence.

## Evidence

The September evidence page retains an unchanged screenshot of the logged-out public profile, its SHA-256 and a structured transcription. The platform is the source of the displayed numbers; the preserved capture supports inspection of the labels and date. The July account-holder capture remains separate. Live values can change and should not be confused with either dated snapshot.

## What is not established

Aggregate counts do not reveal every tested model, prompt, output, adjudication decision or denominator. They do not demonstrate that each counted result represents an independently verified security failure. The public page does not provide a complete reproduction package or private challenge material. It also does not establish paid evaluation work, current ranking or general competence in penetration testing.

## Contribution and next evidence

The relevant contribution is hands-on model-behavior testing and evidence documentation within the demonstrated surfaces. My broader technical implementations are AI-assisted. A shareable method and rubric case could strengthen the record if it were completed and permitted to publish; it is not listed as an existing achievement. For current evaluation-support roles, use the specialist résumé and inspect the dated record alongside the scientific source work and accepted Atlas contributions.

[AI evaluation résumé](/cv-ai)

## Claim provenance

Dated Proving Ground record: rank #74, top 6%, on 29 July 2026.

The displayed total is 113; four visible area counters sum to 112. Arena figures are separate. This is not a current ranking or paid engagement.

Evidence: dated platform snapshot · Owner: platform record captured by account holder · Observed 2026-09-12 · Event/start 2026-07-29.

[Original evidence](https://mariomarcolongo.com/evidence/gray-swan-2026-07-29/)Dated Proving Ground record: rank #79, top 5%, with 127 platform-recorded total breaks on 30 September 2026.

Captured from the logged-out public profile. The displayed total is 127; visible area counters sum to 126. Arena separately shows rank #462, 31 unique breaks, 1,260 points and 354 submissions. This is dated platform evidence, not paid employment, security certification or independent reproduction of every result.

Evidence: dated platform snapshot · Owner: public platform profile · Observed 2026-09-30 · Event/start 2026-09-30.

[Original evidence](https://mariomarcolongo.com/evidence/gray-swan-2026-09-30/)

Source review: 2026-09-30. This is a dated account; the linked sources may have changed.