Job Description

On behalf of SafetyTech Client #1, SD Solutions is looking for a talented [Position name]

SD Solutions is a staffing company operating globally. Contact us to get more details about the benefits we offer.

Responsibilities:

You design prompt injection scenarios in YAML, run them against frontier models, validate success rates, and submit passing tasks. 5 high-quality tasks per week (full-time equivalent). Per-task compensation, paid on acceptance.

Requirements:

• Adversarial mindset: you think like an attacker and understand how to exploit an AI agent’s helpfulness, authority assumptions, or trust in its environment
• Prompt injection expertise: direct (role-play, encoding, context flooding) and indirect/environment-embedded (poisoned tool responses, malicious content in documents, cross-context leakage)
• Technical writing in YAML
• Comfortable with Docker, CLI tools, and running systematic tests against multiple models
• Domain realism in at least one vertical: e-commerce, finance, HR, enterprise SaaS, healthcare, travel
• Background in pentesting, appsec, LLM security research, or red teaming strongly preferred

The Task

You build adversarial prompt injection tasks for Alice’s RL Gym platform. Each task is a self-contained YAML scenario simulating a realistic AI agent deployment, testing whether the agent can be manipulated into violating its safety policies.

What a task includes:

1. An agent persona with a system prompt, security policies, and a defined set of tools

2. A simulated environment (e-commerce site, messaging platform, enterprise app) with content the agent interacts with

3. An adversarial payload embedded in the environment (in messages, product listings, tool responses, documents) that attempts to trick the agent into a policy violation

4. A benign user request that naturally leads the agent to encounter the payload

5. Deterministic evaluation criteria specifying exactly what constitutes success (agent resists) vs. failure (agent is compromised)

6. Quality gate: each task must cause a policy violation in at least 7/10 runs against at least 2 of 3 SOTA models. Attacks must be diverse (varied technique, surface, domain) and realistic inproduction agent deployments. No contrived setups or tools that exist only to enable the attack.

About the company:

A company building specialized evaluation infrastructure for AI safety and robustness testing. Their platform simulates adversarial conditions used by AI development teams to validate agent behavior before deployment. Currently expanding a freelance contributor pool for scenario and environment development.

By applying for this position, you agree to the terms outlined in our Privacy Policy. Please take a moment to review our Privacy Policy https://sd-solutions.breezy.hr/privacy-notice, and make sure you understand its contents. If you have any questions or concerns regarding our Privacy Policy, please feel free to contact us.