Safeguards Enforcement Analyst, Conventional Weapons

Written by the T&S Network from Anthropic's posting.

Enforcement analyst work on Anthropic's AI systems, detecting and mitigating attempts to misuse them to facilitate real-world harm through conventional weapons and dangerous technology. The role builds operational workflows to assess model behavior, drive enforcement decisions, and develop evals, and sits within the Safeguards team.

• Terrorism and violent extremism

What you'd do

• Create automated enforcement systems and review pipelines that run at scale without losing precision
• Write and upkeep evals that gauge model behavior on these policy areas and catch regressions
• Provide the Safeguards policy design group with organized notes on policy holes and unclear enforcement calls
• Flag new misuse trends, unfamiliar attack methods, and indicators of coordinated violence for escalation
• Review flagged content to make enforcement decisions and surface policy gaps, especially novel or sophisticated misuse
• Partner with Engineering and Data Science to tune detection and automated enforcement for policy violations
• Write enforcement guidelines and reviewer documentation for consistent decisions across varied content

What it asks for

• Deep applied expertise in weapons systems, able to translate complex technical evidence into enforcement decisions
• Hands-on time building up and expanding enforcement or content review operations
• Working knowledge of SQL or similar data tools
• Track record spotting emerging risks and threat actors and briefing Product, Policy, Engineering, and Legal
• Hands-on work with generative AI tools, including crafting prompts for review and enforcement
• Grasp of how product rules get applied consistently at scale, including content moderation

Worth knowing

• May be exposed to explicit content that is violent, graphic, hateful, or psychologically disturbing

Terms in this role

OSINT.

Open-source intelligence: investigating with publicly available information, such as social media, public records, archived pages and imagery.

Red teaming.

Testing a system, often an AI model, by deliberately trying to make it fail or misbehave, so problems are found and fixed first.

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on Anthropicnorth_east before you apply. If we got something wrong, report this listing.