We are seeking an AI/LLM Safety Engineer to join our AI team and take ownership of how safely our models and agents behave in production; with a focus on AI Safety, Trust & Safety, and Responsible AI. You will design the evaluations that catch unsafe behavior, build the guardrails that stop it, and lead the red-teaming that finds the gaps before our users—or attackers—do. Agent safety is the primary focus of this role: you will help ensure that as our systems gain the ability to call tools and take actions, they do so within well-defined, well-tested boundaries.

Key Responsibilities:

LLM Safety Evaluation & Red Teaming

• Design and maintain a safety evaluation framework—adversarial prompt sets, scenario-based test suites, and regression suites—so that every model and agent update is validated before it ships.
• Lead structured red-teaming exercises covering jailbreaks, prompt injection, tool misuse, and data exfiltration; document findings and drive each issue through to remediation and closure.

Guardrails & Runtime Controls

• Build and iterate on guardrail logic, including input/output filtering, tool-boundary constraints, action validation, sensitive-data redaction, and policy prompting.
• Integrate safety checks into CI/CD and runtime so that unsafe behavior is intercepted before it reaches users.

Agent Safety (primary focus of this role)

• Perform threat modeling for agentic scenarios: tool-call boundaries, sandbox isolation, and least-privilege access, with particular attention to preventing agents from exfiltrating data or executing irreversible actions through chained tool calls.
• Conduct safety reviews of reinforcement-learning (RL) environments and trajectory data, partnering with environment and agent engineering teams to embed safety constraints directly into the environments themselves.

Monitoring & Observability

• Instrument AI features for safety with structured logging, tracing, and metrics, enabling detection of unsafe patterns and regressions in production.

Governance & Collaboration

• Prepare evidence for governance reviews—test reports, evaluation summaries, and mitigation validation—aligned with internal Responsible AI standards.
• Collaborate with Product and UX to improve safety interactions (warnings, confirmations, refusal messaging, and feedback collection), and align evaluation goals with the Research and Data teams.

• Bachelor's or Master's degree in Computer Science, Software Engineering, Cybersecurity, or a related technical field—or equivalent practical experience.
• 4+ years building production software, with direct experience working on—or securing—ML/LLM systems.
• Strong software engineering skills with the ability to write production-grade code (primarily Python), beyond scripting or notebook prototyping.
• Solid understanding of LLMs and ML: how models work, prompt engineering, and the safety implications of fine-tuning and RAG (e.g., unsafe retrieval, tool misuse, and data exfiltration).
• A security mindset with demonstrated threat-modeling ability; able to threat-model AI workflows and familiar with the fundamentals of access control, data retention, and incident response.
• Familiarity with the LLM attack surface—prompt injection, jailbreaks, data poisoning, and supply-chain risk—and working knowledge of the OWASP LLM Top 10.
• Hands-on experience with at least one of safety evaluation or red teaming, with the ability to walk through a real finding and how it was remediated.

Preferred Qualifications

• Hands-on experience with industry safety tooling such as garak, PyRIT, promptfoo, Giskard, and NeMo Guardrails, and the ability to articulate the trade-offs between them.
• Visible output in AI safety or security: publications at relevant venues (e.g., the NeurIPS AI Safety Workshop, USENIX Security, or DEF CON AI Village), open-source contributions, or responsible disclosures on frontier models with public write-ups.
• Familiarity with AI governance and compliance frameworks (NIST AI RMF, ISO/IEC 42001, EU AI Act) and the ability to translate compliance requirements into concrete engineering tasks.
• Engineering experience with agents, RL environments, and/or tool use.
• Practical experience with threat-modeling methodologies such as MITRE ATLAS and STRIDE/PASTA.

About Propio

Propio is on a mission to make communication accessible to everyone. As a leader in real-time interpretation and multilingual language services, we connect people with the information they need across language, culture, and modality. We are committed to building AI-powered tools that enhance interpreter workflows, automate multilingual insights, and scale communication quality across industries.