About the role

Anthropic is an AI safety and research company working to build reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our customers and society as a whole. As an Enforcement Analyst focused on Agentic Alignment, you will play a crucial role in identifying and investigating harmful and unexpected behavior by models, and make enforcement decisions accordingly. You'll work closely with a suite of cross-functional teams, within and outside of Safeguards, including: Policy, Threat Intel, Engineering, Research, and Transparency. Your work will be essential in maintaining Anthropic's commitment to safe and beneficial AI as Claude takes on more autonomous work.

Important context for this role: In this position you may be exposed to and engage with explicit content spanning a range of topics, including those of a sexual, violent, or psychologically disturbing nature. There is also an on-call responsibility across the Policy and Enforcement teams.

Key responsibilities

• Review and investigate potential instances of Agentic Misalignment, working from transcripts, tool-use traces, and code
• Triage cases by severity and urgency and share through the appropriate channels
• Help operationalize review pipelines for agentic cases, translating leads from detection outputs and investigations into structured review workflows that scale
• Make enforcement decisions under the Usage Policy, applying it consistently to novel agentic scenarios and across surfaces, and documenting the reasoning behind each decision
• Assess cyber-related agentic activity, such as scanning, exploitation attempts, credential misuse or malicious code to determine its real-world harm potential
• Write clear, well-evidenced case write-ups for audiences ranging from enforcement peers to senior leadership
• Identify patterns and trends across cases, and close the loop between enforcement outcomes and upstream improvements by feeding findings back to policy, detection and research teams to improve policies, classifiers and model behavior
• Maintain rigorous documentation of enforcement decisions, review logic and precedents to build institutional knowledge
• Collaborate with Threat Intelligence, Research, Engineering and Policy partners so that detection coverage keeps pace with how agentic misuse and misaligned behavior evolve
• Use AI systems, including Claude, to accelerate your own investigations: summarizing long transcripts, analyzing code, querying data and building lightweight tooling
• Contribute to regular reports on agentic harm trends and enforcement outcomes

Minimum qualifications

• Can read and understand code (for example Python, JavaScript or shell scripts) well enough to follow what happened and why it matters
• Have a background in security, cyber threat intelligence, or a similar technical field that lets you understand how cyber harm happens and judge its severity
• Have experience in trust and safety, enforcement, investigations or a related field, translating ambiguous policy questions and facts into defensible enforcement decisions and clear written findings
• Have conducted structured investigations, building a case from first signal to final finding and knowing when the evidence is strong enough to act on
• Have supported or contributed to formal enforcement actions, including case documentation, evidence packaging and escalation coordination
• Write clearly and concisely, and can present complex findings to stakeholders across Legal, Policy and Engineering, including non-technical audiences
• Have sound judgment about when a case needs escalation, and the resilience to work with difficult content

Preferred qualifications

• Have demonstrated use of AI systems to accelerate your own work, and are excited to keep finding new ways to do so
• Are comfortable working with large volumes of data, from long transcripts, logs and traces to complex, multi-table datasets, and surfacing the behavioral patterns that matter
• Are curious about how AI agents behave, and enjoy investigating behavior no one has categorized before, building approaches where established playbooks don’t exist
• Have familiarity with AI safety and alignment, and current debates about the risks of increasingly autonomous AI systems