Quick answer
AI safety and red-teaming is a genuinely growing field in 2026 — every major AI lab employs people whose job is deliberately trying to break their own models before release, plus policy researchers, alignment researchers, and safety engineers. Backgrounds vary widely: security research, social science, specific domain expertise (extremism, fraud, medicine), and machine learning all lead in, more so than one standard credential.
This field is less standardized than "become a software engineer" — there isn't one degree that gets you in, and that's actually an opportunity for people with unusual, specific expertise that a pure ML background doesn't provide.
The different roles inside "AI safety"
- Red-teamers — deliberately try to make a model fail, using domain expertise or adversarial creativity
- Alignment researchers — work on the technical problem of making AI systems reliably do what's intended
- Policy researchers — work on governance, regulation, and institutional safety questions
- Safety engineers — build the guardrails, monitoring, and evaluation infrastructure labs rely on
Realistic ways in
A security research background transfers directly to red-teaming. Domain expertise (public health, extremism research, financial fraud) is genuinely valued for probing specific harm categories. A machine learning background opens alignment and safety engineering roles. None of these require starting from zero if you already have adjacent expertise — the field actively wants outside perspectives, not just ML PhDs.
If you have deep expertise in a specific harm domain — fraud, extremism, medical misinformation — that expertise is genuinely valuable to an AI safety team, arguably more than a generic ML background alone.
Related reading
Bottom line
AI safety is a real, growing career path with more entry points than people assume — it's not restricted to machine learning specialists. If you have relevant domain expertise or a security background, it's worth a genuine look, not just something for ML PhDs.