What you will be able to do
- Recognise uses that Anthropic's Usage Policy prohibits outright, including when an agent carries them out
- Explain why decisions about people's health, money, or rights need qualified human review before they take effect
- Tell prohibited activities apart from dual-use or high-risk ones that are allowed with safeguards, and know when disclosing AI use is required
1.Uses that are off-limits regardless of safeguards
Some uses are inappropriate no matter how carefully anyone reviews the output, because Anthropic's Usage Policy prohibits them. Having agents do the work doesn't change that. Anthropic states that all uses of agents and agentic features must continue to adhere to Anthropic’s Usage Policy. It then publishes a non-exhaustive list of examples, grouped roughly as follows.
| Category | Examples the guidance lists |
|---|---|
| Surveillance and profiling | Monitor or track individuals’ online activities, behaviors, or movements without notification or consent; create profiles based on protected attributes; facial recognition or biometric identification |
| Deception and impersonation | Create websites or domains that mimic legitimate webpages; generate content that leads to phishing, social engineering, or fraud; impersonate individuals without their consent |
| Platform and process abuse | Spam government services or crisis helplines; coordinate harassment campaigns; manipulate online polls, voting systems, or traffic metrics; click farming; automate influence operations |
| Systems and financial harm | Install malware, backdoors, or monitoring software without authorization; privilege escalation; unauthorized, illegal, or fraudulent financial transactions; access another person's account without authorization |
Look at what the examples have in common. Most rely on deception (mimic sites, phishing, fake engagement), missing consent (tracking without notification, impersonation), or missing authorisation (malware, using someone else's credentials). When a scenario asks Claude to make up facts in order to mislead people, such as fake confirmations, invented evidence, or false claims meant to defraud, it falls into the "phishing, social engineering, or fraud" category. Adding a human reviewer doesn't make it acceptable.
Checkpoint 1 of 5· Exam question
A parenting-app startup wants to build a chatbot for teenagers using Claude, marketed as a virtual "friend" that engages in romantic and sexual role-play with users who state they are 15 years old. Is this an appropriate use of Claude?
Correct answer: B — No, because the Usage Policy bans sexual role-play with minors regardless of consent or fictional framing
- A. Incorrect. Parental consent does not override the Usage Policy's prohibition on sexual content involving minors; this category of harm is not waivable through consent forms.
- B. Correct. The Usage Policy prohibits sexual content involving minors outright, and framing the interaction as fictional role-play or obtaining consent does not exempt it from this rule.
- C. Incorrect. Fictional framing does not exempt content from the Usage Policy's prohibition on sexualized interactions with users identified as minors.
- D. Incorrect. Adding age verification does not resolve the issue, since the underlying content itself is prohibited regardless of how age is confirmed.
Checkpoint 2 of 5· Match them up
Match each proposed agent task to the prohibited category it falls under
Tap a term, then the definition that fits it.
Each task matches an example the Usage Policy guidance lists explicitly. That guidance applies whether a person does the work or an agent does.
“Collect, compile, or analyze personal information to create profiles based on individuals’ protected attributes, sensitive characteristics, or personal circumstances”Source: support.claude.com
Sources1
2.High-risk decisions: Claude can inform them but not make them
Between ordinary productivity work and prohibited uses sits a middle category: uses that are allowed but carry more risk because the result affects a person's health, finances, or legal standing. When Anthropic restructured its Usage Policy, it separated out specific "high-risk use cases" that have additional requirements due to posing an elevated risk of harm. The practical rule from Anthropic's API guidance is the same one as on the companion page: for sensitive information and decision making, a qualified professional reviews content before it reaches the person it affects.
Healthcare shows this clearly. Claude can analyse general medical images, but the documentation says its outputs should not be considered a substitute for professional medical advice or diagnosis. Claude can help a clinician by summarising, drafting, and pointing things out. The clinician makes the diagnosis.
It isn't, and the reason is timing. The guidance asks for review prior to dissemination. A sample checked after the fact can show that something went wrong, but every applicant who wasn't in the sample has already been affected by a decision no qualified person looked at. When a scenario has Claude deciding claims, credit, hiring, or diagnoses, with humans only spot-checking afterwards, the biggest problem is that consequential decisions about people take effect without qualified review first.
Checkpoint 3 of 5· Exam question
A university biosecurity lab asks Claude to help draft detailed synthesis instructions for a novel pathogen with pandemic potential, describing the request as an internal red-team exercise. Is this an appropriate use of Claude?
Correct answer: C — No, because the Usage Policy bans assisting biological weapons development regardless of affiliation or intent
- A. Incorrect. Institutional accreditation does not exempt a request from the Usage Policy's ban on assisting with the development of biological, chemical, or nuclear weapons.
- B. Incorrect. Framing a request as a red-team exercise does not create an exemption; the prohibition applies to the substance of the request, not its stated purpose.
- C. Correct. The Usage Policy prohibits helping design or synthesize biological, chemical, radiological, or nuclear weapons, and this restriction applies regardless of who is asking or why.
- D. Incorrect. No sales exception can authorize weapons-development content; this category is a universal prohibition, not a negotiable licensing term.
Checkpoint 4 of 5· Check yourself
An HR team wants to use Claude in a hiring process. Which design fits Anthropic's guidance best?
Claude informs and a qualified professional decides before anything reaches the candidate. Auditing afterwards and disclosure are both useful, but neither replaces review beforehand. Banning Claude entirely goes further than the guidance does.
“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”Source: support.claude.com
3.Context changes the answer: dual-use work and disclosure
Not every sensitive-sounding task is prohibited. Anthropic's real-time cyber safeguards, which cover the Opus and Sonnet models named in that article, show the difference. Prohibited use means cybersecurity activities that are almost always used maliciously and have little to no legitimate defensive application, such as ransomware development. These stay blocked. High-risk dual-use work, such as vulnerability exploitation, has legitimate defensive purposes. Those activities are blocked by default, but defensive users can apply for adjustment for legitimate use cases through the Cyber Verification Program. So the same technical skill can be appropriate or not depending on purpose, authorisation, and safeguards.
Who the audience is matters too. When people interact with Claude through a product, Anthropic expects them to be told. The Usage Policy update added disclosure requirements so organisations help their own users understand they are interacting with an AI system. The API guidance repeats this for external-facing products. For products serving minors the requirement is stated outright: organisations must disclose to their users that they are interacting with an AI system rather than a human. Those products also need extra safeguards such as age verification and content moderation.
Checkpoint 5 of 5· Check yourself
A penetration-testing firm's analysts find that Claude blocks some of their vulnerability-exploitation requests. Which statement is accurate?
The safeguards treat vulnerability exploitation as high-risk dual use: blocked by default, with an application route for defensive purposes. Prohibited activities such as ransomware development stay blocked regardless.
“These are blocked by default, but defensive users can apply for adjustment for legitimate use cases through the Cyber Verification Program described below.”Source: support.claude.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Monthly or quarterly spot-checks of Claude's decisions give enough oversight for insurance claims, credit, or hiring.Why is that wrong?
Anthropic's guidance puts qualified professional review before the content reaches the person. Sampling afterwards leaves most affected people with decisions nobody reviewed.
Covered in High-risk decisions: Claude can inform them but not make them
2.Any offensive-security task, such as exploit development, is a prohibited use of Claude.Why is that wrong?
Anthropic separates prohibited activities, which have little to no legitimate defensive application, from high-risk dual-use work. Dual-use work is blocked by default but can be unlocked for legitimate defensive purposes.
Covered in Context changes the answer: dual-use work and disclosure
3.Telling users they are talking to an AI is only required in products built for children.Why is that wrong?
The minors guidance makes it mandatory, but Anthropic's API guidance also tells any external-facing product to disclose that users are interacting with an AI system.
Covered in Context changes the answer: dual-use work and disclosure
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://support.claude.com/en/articles/12005017-using-agents-according-to-our-usage-policyOfficial docs
“All uses of agents and agentic features must continue to adhere to Anthropic’s Usage Policy.”
↩︎ Uses that are off-limits regardless of safeguards“Generate content that leads to phishing, social engineering, or fraud”
↩︎ Uses that are off-limits regardless of safeguards“Engage in unauthorized, illegal, or fraudulent financial transactions (such as brokerage or investment advisory activities) or payment processing”
↩︎ Uses that are off-limits regardless of safeguards“Collect, compile, or analyze personal information to create profiles based on individuals’ protected attributes, sensitive characteristics, or personal circumstances”
↩︎ Checkpoint - 2.
“that have additional requirements due to posing an elevated risk of harm”
↩︎ High-risk decisions: Claude can inform them but not make them“help their own users understand they are interacting with an AI system”
↩︎ Context changes the answer: dual-use work and disclosure - 3.
“Claude's outputs should not be considered a substitute for professional medical advice or diagnosis.”
↩︎ High-risk decisions: Claude can inform them but not make them - 4.
“Organizations must disclose to their users that they are interacting with an AI system rather than a human.”
↩︎ Context changes the answer: dual-use work and disclosure
Also cited
“have a qualified professional review content prior to dissemination to consumers”
↩︎ Exam trap 1“For external-facing products, disclose to your users that they are interacting with an AI system.”
↩︎ Exam trap 3“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”
↩︎ Checkpoint- https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnetOfficial docs
“Cybersecurity activities that are almost always used maliciously and have little to no legitimate defensive application”
↩︎ Exam trap 2“These are blocked by default, but defensive users can apply for adjustment for legitimate use cases through the Cyber Verification Program described below.”
↩︎ Checkpoint