[EJV] Penetration Tester — AI Safety & Security

We are looking for a Penetration Tester to join our AI Safety & Security team, working hands-on across Bosch's growing portfolio of AI-enabled products and internal tooling — LLM deployments, agentic AI systems, and Model Context Protocol – style tool-calling integrations. Under the direction of the team's Lead AI Safety & Security Engineer, you will execute structured test plans against real systems, covering both whether an AI system behaves safely (does it refuse harmful requests, stay within its intended scope) and whether it is secure (can it be manipulated via prompt injection or tricked into misusing its tools). This is an individual-contributor role on a newly forming team — you'll be running real engagements from day one, not just shadowing them.

• Execute penetration tests and safety evaluations against LLM-based features and AI agents, following test plans and specifications defined by the team lead.
• Run red-teaming and jailbreak testing to surface harmful, biased, or unsafe model outputs, and document reproducible findings.
• Test for prompt injection (direct and indirect), excessive agency, and tool/plugin abuse in agentic and MCP-style tool-calling systems.
• Verify human-in-the-loop and guardrail controls hold up under adversarial conditions, not just in normal use.
• Write clear, reproducible findings reports — evidence, severity, and a specific recommended fix — that both engineers and governance stakeholders can act on.
• Help build and maintain test tooling: adversarial-prompt scripts, jailbreak test cases, and MCP/tool-schema fuzzing scripts.
• Track published AI security research (OWASP LLM and Agentic Top 10s, MITRE ATLAS) and bring new attack techniques into the team's test suite.
• Work directly with AI/ML engineering teams to reproduce, triage, and confirm fixes for reported findings.
• Support the team lead in scoping and coordinating external red-teaming/pentest vendor engagements when work is outsourced.