Anthropic’s safety hiring looks like a threat map

Anthropic’s latest hiring spree reads less like a careers page and more like a risk register.

The company is bringing in enforcement analysts whose job is to stop its models from teaching people how to build nuclear, chemical, and biological weapons, run scams, or commit cybercrime, as first reported by Axios.

Job ads as threat assessment

Most companies sell a mission. Anthropic’s listings quietly list what it fears most.

One opening seeks an “Enforcement Analyst focused on Radiological & Nuclear Harms.” Others explicitly cover chemicals and explosives, financial fraud, and related abuse areas.

Recommended reading

Mira Murati’s lab unveils Inkling, a 975B open model

The roles pay in the mid- to upper-$200,000s and are not traditional engineering jobs. Anthropic is looking for real-world expertise in fields like biology and explosives, plus people who can think like an attacker trying to slip past its defenses.

A spokesperson said the blunt language in the job titles is intentional:

“Ensuring our models don’t provide potentially harmful information is central to responsible development.”

The company said it regularly hires experts in sensitive fields to stress-test its models before a release, and that spelling out the exact harm is how it reaches the right candidates.

Anthropic says hundreds of staff now work on safety, probing for weak spots and patching them.

Amodei’s catastrophe scenario

The hiring pattern doubles as a response to critics who call Anthropic the industry’s biggest doomsayer. The company is now spending heavily on the very risks it has been talking about.

Chief executive Dario Amodei has been publicly outlining his worst-case views for months.

In a January essay he wrote that biological attacks are the scenario that worries him most:

“I do not think biological attacks will necessarily be carried out the instant it becomes widely possible. But added up across millions of people and a few years of time, I think there is a serious risk of a major attack, with casualties potentially in the millions or more.”

Amodei has also warned about models helping cybercriminals and empowering authoritarian states.

Earlier this year, Anthropic broke with the US Defense Department over the use of its technology for mass surveillance and autonomous weapons.

Labs write their own rules

Anthropic is not alone in trying to operationalize worst-case thinking.

OpenAI is hiring a researcher on biological and chemical risks, offering a base salary of up to $445,000.

As models become more capable, “every serious lab is racing to staff a red team,” the article notes. But that race is happening without a legal framework.

The US still has no comprehensive AI safety law. Congress has tried for years and passed nothing.

Some industry leaders want an external referee. Google’s Demis Hassabis has floated a Wall Street-style watchdog for frontier models.

For now, the expertise remains heavily concentrated in the private sector: fewer than one in a hundred AI PhDs go into government, so the people who understand the systems best mostly work inside the labs themselves.

That produces a tight loop of self-regulation. The same firms building what Amodei has called the “most dangerous capability” are also deciding how to fence it in.

Amodei has named that tension himself, describing AI companies as the next tier of risk after hostile states. Anthropic’s careers page ends up serving as both argument and warning: the people best placed to stop the catastrophe are being hired by the company that could help cause it.