AI Safety, Evals and Governance Jobs: 91 of 2,852 Open AI Roles (September 2026)

The three AI functions people ask us about most are the three with the least inventory. On 7 September 2026 the board held 2,852 live AI roles, read directly from the applicant tracking systems of the companies we track. Of those, 47 were classified AI Safety & Alignment, 34 Evals & Red Teaming and 10 AI Governance & Policy. That is 91 roles, or 3.2% of the market, roughly one open AI role in 31. Applied AI alone held 955.

This is the first monthly edition of the scarcity index. Every figure is frozen at publication, the method is at the bottom, and the live numbers are always at /api/public/stats for anyone who wants to check.

The nine functions, ranked

We classify every role into one of nine functions from its title and description. Here is the whole board on the day of the read:

• Applied AI - 955
• ML Engineering - 505
• Research Science - 492
• Data & MLOps - 403
• AI Infrastructure - 348
• AI Product - 58
• AI Safety & Alignment - 47
• Evals & Red Teaming - 34
• AI Governance & Policy - 10

The bottom three are not a rounding error in a large market. They are a different market, with a handful of employers, almost no remote work, and unusually high pay disclosure.

AI safety: one employer is 70% of the function

The 47 safety and alignment roles sat at just 10 companies. Anthropic held 33, OpenAI 6, and Character.AI, ElevenLabs, Hark, Isomorphic Labs, Lila Sciences, Northrop Grumman, Roblox and Waymo one each. Remove one employer and the function shrinks by 70%. The live company list is at companies hiring AI safety researchers.

A stricter cut, roles with "safety" or "alignment" in the title, returns 18 on the AI safety jobs hub. Those 18 tell you what the work looks like: 13 of them publish a pay band, none is fully remote, and 8 are hybrid. The disclosed bands run from Hark's AI Safety Engineer at $150,000 to $300,000, through Lila Sciences' Staff / Principal Research Engineer at $224,000 to $336,000 and Roblox's Senior Engineering Manager, AI Safety Platform at $345,040 to $399,420, to OpenAI's Researcher, Alignment Interpretability at $295K to $500K and Anthropic's Staff+ Software Engineer, Infrastructure, Interpretability at $405,000 to $485,000. Waymo's London safety researcher role is the one non-US band, at £113,000 to £119,000.

Read those titles again. Half of them are engineering roles: infrastructure for interpretability, safety platform management, security engineering with a safety remit. The research scientist seat with a publication requirement is the minority of the function, not the whole of it.

Evals: 23 employers, and a platform company at the top

Evals & Red Teaming is the most spread out of the three: 34 roles across 23 companies. The largest single employer was LangChain with 8, every one of them for its evals and observability platform rather than for evaluating a model of its own. Anthropic, Cohere, Mercor and OpenAI held 2 each. The rest, including METR, Apollo Research, Artificial Analysis, Cartesia, Deepgram, Exa, Fireworks AI, Glean, Krafton and Legora, held one. Full list at companies hiring AI evaluation engineers.

The model evaluation hub, which matches on the title and so also catches eval-titled roles filed under other functions, listed 39. Pay disclosure there is the highest we see anywhere on the board: 29 of the 39 publish a band. The spread is wide. LangChain's Frontend Engineer, AI Observability & Evals Platform is $160,000 to $195,000; Reducto's Machine Learning Eval Engineer is $200K to $300K; Cohere's evaluation roles in London and Toronto are CA$250K to CA$535K; OpenAI's Agent Post-Training, Frontier Evals and Environments Research is $380K to $500K; and Anthropic's Research Engineer, Model Evaluations tops the function at $500,000 to $850,000.

Two of the 39 are remote. And exactly one live role on the whole board carries a red-team title, Apollo Research's AI Red Team Engineer in London and San Francisco, which is the entire content of the AI red team hub today.

Governance: ten roles, nine companies, and a lot of interest

Ten roles were classified AI Governance & Policy on the day of the read. On the same day the companies-hiring page, which is cached daily, showed nine companies with exactly one role each: AlphaSense, Anthropic, Capital One, Cohere, Databricks, KLA, Scale AI, Upstart and xAI. Nobody is building a governance team; nine companies are each hiring one person.

The title hubs split the function further: 7 governance-titled, 2 policy-titled and 2 compliance-titled roles. Two of the ten publish pay: Upstart's Senior Manager, Loan Salability & First-Line Model Governance at $172,100 to $238,300 (and remote, the function's only remote role) and Databricks' Principal Product Marketing Manager, AI Governance at $154,800 to $212,850. The oldest open governance role was posted in January.

The demand side is the opposite shape. In late August a single LinkedIn post about the governance hub, reshared once by an AI governance practitioner, sent about 1,450 clicks to a page that held ten roles at the time. We saw the same spike in our own site analytics. The interest in this function is enormous relative to its inventory, and anyone planning a move into it should plan around that ratio, not around the headlines about AI regulation creating jobs. The regulation is real; the EU AI Act alone is referenced by dozens of live roles across other functions. The dedicated seats are not there yet.

Remote work barely exists in these three functions

Of the 378 fully remote AI roles on the board that day, zero were AI Safety & Alignment, one was Evals & Red Teaming and one was AI Governance & Policy. Remote AI work lives in Applied AI (116 remote roles), ML Engineering (97) and Data & MLOps (81). If a remote safety role is the goal, the honest reading of the data is that there is not one to apply to today.

What to do with this if you want in

• Apply to the function by its engineering door. Most safety and evals roles that publish a band are infrastructure, platform and evaluation engineering seats, not research scientist seats. A public eval harness with real results, or a reproduced interpretability paper with the gaps written up, is the artifact those screens look for.
• Treat governance as a one-seat-per-company market. Nine companies, one role each. The neighbouring functions with real volume are privacy, model risk management in banking, and security GRC, which is where several of these ten postings sit organisationally.
• Expect to relocate or go hybrid. Two remote roles across 91 is the number.
• Watch the company lists, not the hubs. With 10 employers in safety and 23 in evals, the practical move is to read those companies' feeds directly. The pages linked above refresh hourly.

Method and caveats

Counts are live roles on 7 September 2026, read hourly from each company's own applicant tracking system (Greenhouse, Lever, Ashby, Workday and Comeet) and classified into nine functions by title and description keywords. It is a large sample of the companies that publish a feed, not a census: companies with no public board are invisible to us, so treat every count as a floor. Two universes appear above and are labelled each time: the classifier count (47 / 34 / 10) and the title-hub count (18 / 39 / 7), which matches on the job title only and can include roles the classifier files elsewhere. Company lists come from pages cached daily, which is why the governance list shows nine roles against a live count of ten. Pay bands are quoted exactly as the employer published them, in the employer's currency. Every figure here is free to quote with attribution to AI Hiring Board.