About Anthropic
Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.
We’re building a team that will research and mitigate extreme risks from future models.
This team will intensively red-team models to test the most significant risks they might be capable of in the area of Cybersecurity. We believe that clear demonstrations can significantly advance technical research and mitigations, as well as identify effective policy interventions to promote and incentivize safety.
As part of this team, you will lead research to baseline current models and test whether future frontier capabilities could cause significant harm. Day-to-day, you may decide you need to finetune a model to see whether it becomes superhuman in an eval you’ve designed; whiteboard a threat model with a national security expert; test a new training procedure or how a model uses a tool; or brief government, labs, and other research teams. Our goal is to see the frontier before we get there.
We’re currently hiring for our Cyber workstream (as outlined in our Responsible Scaling Policy). By nature, this team will be an unusual combination of backgrounds. We are particularly looking for people with experience in these domains:
Cybersecurity: You’re a white hat hacker who is curious about how LLMs might be able to do some or all of your work. You’re an academic who researches RL for cybersecurity. You’ve participated in or built CTF competitions and cyber ranges, and you want to automate them.
Evaluations: You’ve managed a large-scale benchmark development project, in AI or other domains. You have ideas about how AI and ML evaluations can be more realistic. You’ve thought hard about all the ways to improve model performance on a given benchmark.
Do not rule yourself out if you do not fit one of those categories - it’s plausible the people we’re looking for do not fit any of the above! If you think about the most significant upsides and downsides of AI, and you can do good research to get glimpses of what those look like, please consider applying
Responsibilities
- Design, run, and analyze scientific experiments to advance our understanding of large language models and their application to cybersecurity tasks
- Lead technical design discussions to ensure our infrastructure can support both current needs and future research directions
- Collaborate with other engineers to maintain our evaluations codebase
- Work with external partners to develop novel evaluations to accurately assess the cybersecurity implications of our models
- Partner closely with researchers, data scientists, policy experts, and other cross-functional partners to advance Anthropic’s safety mission
You may be a good fit if you
- Have strong software engineering or AI/ML research experience, and strong interest or experience in offensive cybersecurity
- Have a strong interest in societal and policy impacts of AI
- Take pride in writing clean, well-documented code in Python that others can build upon
- Have a track record of using technical infrastructure to interface effectively with machine learning models
- Have familiarity with prompting and engineering large language models
- Are able to balance research goals with practical engineering constraints
- Have strong problem-solving skills and a results-oriented mindset
- Have excellent communication skills and ability to work in a collaborative environment
- Prefer fast-moving collaborative projects to extensive solo efforts
Strong candidates may also have experience with
- Training, scaffolding, or evaluating models for cyber capabilities
- Competition CTF challenges
- Professional cyber pentesting
- Building sophisticated and realistic cyber range environments
- Representative projects applying large language models or machine learning to cybersecurity tasks
- Developing evaluations or benchmarks for large language models
Candidates need not have
- Previous professional experience in AI safety
- 100% of the skills needed to perform the job
