AI Guides › Playbooks
By Nigel Guy · 7 min read
When a headline says an AI company is hiring someone to police weapons, most people do one of two things: they file it under "not my problem" or they treat it as a sign the end is near. Both reactions feel reasonable and both leave you exactly where you were: building work on Claude without knowing which of your tasks sit near a line Anthropic is actively redrawing. If you're in engineering, drones, robotics, security, defence supply chains, or any field where the same knowledge can build or harm, the people being hired are writing the rules your prompts will be judged against.
The rule: treat Anthropic's policy hiring as advance notice of where enforcement will tighten, and map your own use cases against the published Usage Policy now, while nothing is broken.
Here is the part that can be verified at the time of writing:
What I couldn't confirm: whether the weapons role has been filled, or who took it. Coverage calls it a "hire"; the primary source is a job advert. If you see a name attached to it, check that the name came from Anthropic.
The job description itself gives the reason: conventional weapons are increasingly defined by software, and models can now write guidance code or operate systems. The usage rules are not starting from nothing. Anthropic's Usage Policy already bans using its products to "produce, modify, design, or illegally acquire weapons" and to design weaponisation and delivery processes. What a dedicated owner changes is the precision of that line. A broad ban is applied by judgement; a specialist builds classifiers, threat models and escalation routes. That usually means more consistent enforcement, and some legitimate edge-case work getting flagged until the boundaries settle.
The Usage Policy also says plainly that Anthropic will update it "as our technology and the associated risks evolve", that its Safeguards team runs detection and monitoring, and that breaches can lead to throttling, suspension or termination, or to outputs being blocked or modified. So the hiring is the people side of a process the policy already describes.
| You | Why it matters to you |
|---|---|
| Engineers in aerospace, robotics, guidance, sensors, materials | Your normal work is exactly the "dual-use" zone the role is built around |
| Security and pen-testing teams | The policy bans finding or exploiting vulnerabilities without the system owner's permission; you need a clear authorisation trail |
| Businesses building on the Claude API | Your customers' prompts count as your inputs; an enforcement action can stop your product |
| Anyone in the policy's High-Risk Use Cases (legal, healthcare, insurance, finance, employment and housing, academic testing, published media) | Separate duties already apply: a qualified human reviews before anything reaches a person, and you disclose AI use |
| Everyone else writing emails and summaries | Very little changes. Read the What to skip section |
The Boundary Map is a one-page table you keep for every recurring way your team uses Claude. It has five columns.
Then run it on a cycle:
Fill in the bracketed parts and paste the current Usage Policy text where shown.
You are a careful compliance analyst helping a small business understand how its
use of Claude relates to Anthropic's published Usage Policy. You are not giving
legal advice.
Context:
- Our business: [ONE_SENTENCE_DESCRIPTION_OF_BUSINESS]
- Our sector and country: [SECTOR], [COUNTRY]
- How we use Claude, one use case per line: [LIST_OF_USE_CASES]
- The current Usage Policy text, copied today: [PASTE_POLICY_TEXT]
Goal: produce a Boundary Map so we know which use cases sit near a policy line.
Steps:
1. For each use case, find the single closest clause in the pasted policy text
and quote its heading exactly. If nothing applies, write "none".
2. Score each use case Clear, Adjacent or Contested, with a one-sentence reason.
3. For Adjacent and Contested rows, list the evidence that would show the purpose
is legitimate (for example a contract or written authorisation).
4. Suggest one practical fallback for each Contested row.
Output: a table with columns Use case | Closest clause | Distance | Reason |
Evidence needed | Fallback. Then a short list of questions you could not answer.
Rules:
- Use only the pasted policy text. Do not rely on memory of older versions.
- If a use case is too vague to score, ask me about it instead of guessing.
- Do not suggest ways to phrase requests so they avoid detection.
Before answering, check: does every clause you quoted appear word for word in the
pasted text, and has every use case been scored?
A five-person UK firm makes inspection drones for wind farms. They use Claude for four things: writing customer emails, summarising flight logs, drafting firmware test plans, and researching how payload mounts are certified.
Their map scores emails and log summaries Clear. Firmware test plans come out Adjacent: the subject is flight control software, the purpose is civil inspection, and the evidence is their customer contracts and aviation authority paperwork. Payload mount research comes out Contested, because questions about mounting and releasing payloads read very differently without context. Their fallback for that row is a named external certification consultant. They now include the civil purpose and the relevant contract reference in the project instructions for that work, and they review that row first whenever the policy page changes.
Nothing about their work changed. What changed is that a refusal or a warning email would now be an expected, planned-for event rather than a surprise.