Hand-checked snapshot · 30 September 2026

About one in six AI safety vacancies are open to someone starting out

We read every one of the 141 vacancies currently posted by 15 AI safety organizations, taken straight from their own hiring systems. By the postings' own words, 25 are open to someone with two years of experience or less. 93 are not. The other 23 don't say.

15 standing "expression of interest" listings are left out, since they invite people to register rather than fill a job. Anthropic is kept separate as a comparison.

"Open" is a deliberately modest bar. It means the posting itself makes room for someone early in their career: an intern, associate or junior level, a stated floor of two years or less, or a plain statement that junior or self-taught applicants are welcome. It does not mean beginners actually get hired. If every "doesn't say" posting turned out to be open too, the share would reach 34%, so that is the ceiling. The method page gives the exact rules.

The way in is mostly in the small print

Only 6 of the 25 open roles say so in the title. The other 19 say it in the body of the posting, often in a single sentence:

“We are open to hires at junior, senior, staff and principal research scientist / engineer levels.”UK AI Security Institute, Alignment Red Team - Research Engineer/Research Scientist

“We don’t require a formal background or industry experience and welcome self-taught candidates.”Apollo Research, Research Scientist/Engineer (Evaluations)

“We encourage applications from strong software engineers who are new to ML, and from academics without industrial experience in software engineering.”FAR.AI, Research Engineer

So anyone skimming titles sees a much steeper cliff than the real one. That includes any analysis that counts title keywords, as this site did until September 2026. Titles alone flag 4% of these vacancies as early-career, against the 18% found by reading them. If you are early in your career, read the requirements, not the title.

Which kinds of work have a way in

Among the larger groups, research has the widest door (24% open) and go-to-market the narrowest (11%). Split more simply, 20% of technical roles are open against 14% of non-technical ones. Research, often assumed to be the hardest door, is open 24% of the time, above the overall rate, largely because several labs state plainly that they hire researchers at every level.

Which organizations have a way in

9 of the 15 organizations have at least one open role, and 64% of all open roles sit at four of them: UK AI Security Institute (6), Apollo Research (4), FAR.AI (3), Kairos (3). The biggest boards are not the most open ones, so the field-wide share depends heavily on which organizations are counted. That is why the next check matters.

Does the finding survive a different boundary?

Which organizations count as "AI safety" is a judgment call, so here is the same measure under other reasonable boundaries. The last column runs from the open share to the ceiling, where every "doesn't say" is counted as open.

Across the boundaries of the first five rows the open share stays between 15% and 20%. The finding does not hinge on one organization or one definition.

A frontier lab, for comparison

Anthropic's public board is far larger (636 vacancies) and mostly not safety work, so it is kept out of every figure above. Measured by the same rules, since it has not been hand-checked, 1% of its vacancies are open to early-career applicants (7 roles, 3 of them Fellows Program tracks), against 17% at the AI safety organizations. The AI safety organizations are, if anything, easier to get into than the lab.

What this can and cannot tell you

• Postings are not the whole market. Much senior hiring, and some junior hiring, happens through networks and never gets posted. This measures advertised demand, not where the field most needs people.
• Programs are mostly outside the frame. Fellowships such as MATS, SPAR and GovAI's seasonal fellowships are the main early-career route into the field, and most run on their own application cycles rather than through a hiring system. When one appears as a posting it is counted; the rest are not.
• "Open" is what the posting permits, not who gets hired. A role open to juniors may still go to someone senior.
• Coverage is partial. 15 organizations with a readable hiring system are included. The coverage page lists everyone considered and why each is in or out.
• One reading. The labels come from one careful pass through each posting. Every label and the line behind it is in the audit file, so you can disagree with any of them.