Haris U.

Key Skills

Software

Don't disclose

Other

Top Subject Matter

LLM conversational agent security testing and safety evaluation

Adversarial evaluation of LLMs

multimodal models

Top Data Types

Text

Document

Image

Top Task Types

Red Teaming

Data Collection

Classification

Freelancer Overview

MasterMind (Invite-Only) — Mozilla 0din Generative AI Bug Bounty Program. Brings 6+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Don't disclose and Other. Education includes Bachelor of Science, The Islamia University of Bahawalpur (2024). AI-training focus includes data types such as Text, Document, and Image and labeling workflows including Red Teaming, Data Collection, and Classification.

Labeling Experience

AI Red Teamer - Mozilla

ImageClassificationRed Teaming

Serves as a red teamer in Mozilla’s 0din Generative AI Bug Bounty Program. Performs adversarial testing of conversational models by discovering and documenting prompt injection, jailbreak, and multi-turn manipulation vulnerabilities. Produces reproducible attack cases with severity classification and remediation guidance using structured methodologies. • Authored 41 validated vulnerability submissions and achieved a top global rank • Investigated prompt injection and universal jailbreak weaknesses in production models • Wrote reproducible attack cases with methodology and risk framing • Contributed to systemic safety insights across mainstream generative AI systems

2025 - Present

Researcher (Invite-Only) — Anthropic Model Safety Bug Bounty Program

Don't discloseTextRed Teaming

Participated in an invite-only red teaming program focused on probing and stress-testing Claude model safety boundaries. Conducted adversarial probing to surface misuse patterns and assess robustness against policy and safety constraints. Generated safety test findings as inputs for model safety improvement efforts. • Performed adversarial probing and misuse case discovery • Tested safety boundaries using crafted adversarial prompts • Evaluated model responses for policy and safety compliance failures • Contributed to structured safety validation activities within the program

2025 - Present

MasterMind (Invite-Only) — Mozilla 0din Generative AI Bug Bounty Program

Don't discloseTextRed Teaming

Performed adversarial testing and systematic evaluation of conversational AI systems by authoring reproducible jailbreak and prompt injection attack cases for a public generative AI bug bounty program. Documented vulnerability details with severity classification and remediation guidance to support safety validation. Ranked among the top contributors and produced validated submissions that acted as structured safety test inputs for model evaluation. • Crafted multi-turn manipulation and universal jailbreak scenarios • Authored structured vulnerability reports with reproducibility steps • Tested production conversational models including ChatGPT and Gemini • Supported risk identification across mainstream generative AI systems

2025 - Present

Red Teamer - GraySwan AI

ImageRed Teaming

Competes and researches adversarial model attacks in GraySwan AI’s red teaming arena and invite-only program. Achieves top leaderboard placements by identifying hidden behaviors and weaknesses across conversational, visual, and agentic environments. Translates findings into clear, actionable reports for closed engagements and client programs. • Ranked #28 globally with 939+ successful model breaks in the public arena • Earned $6,600+ through competitive red teaming challenges • Delivered findings across multiple closed client engagements • Produced research outcomes targeting chain-of-thought and multimodal/agent vulnerabilities

2024 - Present

Red Teamer — GraySwan AI (Public Arena + Private Invite-Only Program)

Don't discloseTextRed Teaming

Conducted competitive and invite-only red teaming against AI models to discover misuse cases and elicit harmful or constrained outputs. Produced challenge findings across multiple environments, including text, visual/multimodal, and agentic tool-use settings. Achieved high leaderboard placements and earned bounty rewards based on successful model breaks. • Delivered findings for Hidden Chain-of-Thought and harmful assistant challenges • Exploited multimodal and agentic vulnerabilities through engineered prompts • Participated in closed client engagements under an invite-only program • Documented attack outcomes to demonstrate safety boundary failures

2024 - Present

Show more

Education

Work History