Description
Salesforce’s Office of Ethical and Humane Use is seeking an experienced responsible AI data scientist with an adversarial approach and experience conducting ethical red teaming to contribute to our ethical red teaming practice. In this role, you will help us gain a deep understanding of how our models and products may be leveraged by malign actors or through unanticipated use to cause harm. In addition to adversarial testing, you will analyze current safety trends, and develop solutions to detect and mitigate risk, while working cross-functionally with security, engineering, data science, and AI Research teams. You will bring technical depth to the assessment of AI products, models, and applications, in order to identify the best technical mitigations to identified risks.
The ideal candidate will have technical experience in generative as well as predictive artificial intelligence and in responsible / ethical AI
Responsibilities:
- Adversarial Testing
- Provide technical leadership in designing, prototyping, and implementing comprehensive adversarial testing strategies, including both automated and manual adversarial testing approaches
- Mentor and guide collaborator teams on adversarial testing standard processes, helping them develop the skills to conduct their own testing effectively
- Collaborate with cross-functional teams to integrate OEHU adversarial testing frameworks into the AI development lifecycle
- Safety and Robustness
- Contribute to the development of detection models, safety guardrails, and other proactive measures to prevent and mitigate risks posed by bad actors
- Research and implement innovative techniques for enhancing AI safety and robustness, drawing from both open-source and internal tools
- Collaborate with Salesforce’s AI Research team on novel approaches to model safety
- Technical Research and Implementation
- Write clean, efficient, and well-documented code (primarily in Python) to support research efforts and facilitate the evaluation of AI systems
- Develop and maintain a repository of reusable code modules and libraries to streamline adversarial testing processes
- Testing Execution and Collaboration
- Participate in scoping, documenting, and implementing tests with partner teams, including the implementation of mitigations identified during testing
- Test for technical vulnerabilities, model vulnerabilities, and harm/abuse including but not limited to bias, toxicity, and inaccuracy
- Participate in labeling test data in partnership with OEHU and partner teams
- Reporting, Documentation, and Continuous Learning
- Write reports covering the goals and outcomes of testing operations, including significant observations and recommendations
- Continuously monitor and analyze emerging threats and vulnerabilities to inform the development of adaptive safety measures
- Continue to grow expertise in model safety by keeping up with research in socio-technical systems, privacy, interpretability/explainability, robustness, alignment, and responsible AI
Qualifications:
- 3-5 years of industry experience in Software Engineering, AI ethics, AI research, Applied research, ML, DS, or similar roles
- Demonstrated ability to think adversarially, ability to anticipate how malicious actors might misuse AI systems and develop corresponding test scenarios
- Experience creating heuristic-based detection logic and rules for identifying anomalous or suspicious activity in production systems and networks (e.g. log analysis, user behavior analytics)
- Experience with problem-solving and troubleshooting sophisticated issues with an emphasis on root-cause analysis
- Experience in analyzing sophisticated, large-scale data sets and communicating findings to technical and non-technical audiences
- Proven organizational and execution skills within a fast-paced, multi-stakeholder environment
- Experience working in a technical environment with a broad, cross-functional team to get results, define requirements, coordinate assets from other groups (design, legal, etc.), and deliver key achievements
- Excellent written and oral communication skills, as well as social skills, including the ability to articulate technical concepts to both technical and non-technical audiences.
- Works well under pressure, and is comfortable working in a fast-paced, ever-changing environment
- Experience using SQL and relational databases. Ability to use Python, R, or other scripting languages to perform data analysis at scale
- A related technical degree required
Preferred Qualifications:
- Industry experience in AI red-teaming, implemented manual, or semi-automated, or automated adversarial testing approach in production system
- Experienced with specific AI attack methodologies (prompt injection, model extraction, adversarial examples, etc.)
In office expectations are 36 days/a quarter to support customers and/or collaborate with their teams.
For roles in San Francisco and Los Angeles: Pursuant to the San Francisco Fair Chance Ordinance and the Los Angeles Fair Chance Initiative for Hiring, Salesforce will consider for employment qualified applicants with arrest and conviction records.
