The opportunity
The Intelligence and Investigations team is dedicated to ensuring the safe, responsible deployment of AI by rapidly detecting and mitigating abuse. Our team leverages the latest testing methodologies to uncover vulnerabilities and emerging threats, helping safeguard OpenAI’s products and users.
What you'll do
Design and run rigorous evaluations of model cyber capabilities and: safeguards, including policy adherence, correct refusal, over refusal, and resilience to jailbreaking and other adversarial techniques.
Conduct hands-on testing to understand what models can enable when used by: experienced security practitioners, including through task-specific harnesses, scaffolding, and multi-step workflows.
Distinguish benchmark or policy failures from behavior that creates: meaningful real-world risk, considering feasibility, attacker uplift, reliability, and the capabilities already available elsewhere.
Build and improve automated testing infrastructure that supports repeatable: measurement, rapid iteration, and statistically grounded analysis across models and product surfaces.
Test novel abuse risks in agentic systems, including indirect prompt: injection, agent hijacking, and other ways adversaries may manipulate systems that use tools or act on external information.
Translate findings into clear risk assessments and actionable recommendations: for Security, Research, Product, Policy, and Engineering partners.
What they're looking for
- Contribute a portion of your time to Safety Bug Bounty work, particularly: where reports require cyber expertise.
- Cybersecurity, such as application security, penetration testing,: vulnerability research, adversary simulation, or red-team operations. You can assess whether a demonstrated capability is reliable, feasible, and meaningfully dangerous.
- AI model evaluation, such as designing and running evals, building agentic: harnesses, automating adversarial testing, constructing datasets, or analyzing model behavior at scale. You can turn an ambiguous risk question into a reproducible testing approach.
- Working literacy across both cybersecurity and model evaluation, with an: interest in developing further depth outside your primary area.
