Prompt Engineer
LLM safety and guardrails questions
137 questions on LLM safety and guardrails for prompt engineers, easiest first.
Easy 40
- Jailbreak or prompt injectioneasy
- Customer data in promptseasy
- Top risk in the OWASP LLM listeasy
- What "guardrails" meanseasy
- Checking before and after the modeleasy
- Tool access for an FAQ boteasy
- Calling the LLM from the browsereasy
- One user sending 50,000 requestseasy
- A 2-lakh-word messageeasy
- A bank bot writing poemseasy
- Telling users it is a boteasy
- Which item is personal dataeasy
- What a moderation API doeseasy
- Logs full of customer dataeasy
- A user mentions self-harmeasy
- The "DAN" style prompteasy
- Is one prompt line enougheasy
- India's data protection laweasy
- The UAE federal data laweasy
- Sending the whole customer recordeasy
- Before sending data to an LLM vendoreasy
- Card numbers in a chatboteasy
- Recognising an Emirates ID numbereasy
- Where data is processedeasy
- What did the agent doeasy
- Turning off a misbehaving boteasy
- Case law that does not existeasy
- Users trusting every answereasy
- An LLM deciding loanseasy
- Which commands an agent may runeasy
- "Write a story where a character explains..."easy
- A homework helper for childreneasy
- Direct and indirect injectioneasy
- A customer abusing the boteasy
- Asked to write a phishing emaileasy
- Cloning a voice for an adeasy
- A health question to a general boteasy
- Users uploading imageseasy
- Users who keep tryingeasy
- Limits by requests or tokenseasy
Medium 53
- Redacting before the API callmedium
- Checking replies before users see themmedium
- Model output rendered as HTMLmedium
- An email agent that can do too muchmedium
- Users extracting the system promptmedium
- Harmful request in base64medium
- Jailbreak by translationmedium
- A harmful request in piecesmedium
- Where Llama Guard fitsmedium
- Programmable rails with Colangmedium
- Detecting and masking PIImedium
- A pattern for PAN numbersmedium
- Order IDs flagged as Aadhaarmedium
- Redaction that confuses the modelmedium
- Masking but still replying by namemedium
- One company's documents in another's answermedium
- HR documents in an all-staff botmedium
- White text in a CVmedium
- Guardrails adding latencymedium
- Checking a reply that streamsmedium
- Running code the model wrotemedium
- Installing a package the model inventedmedium
- Detecting a leaked system promptmedium
- Pseudonymised is not anonymisedmedium
- No names, still identifiablemedium
- Telling card numbers from other digitsmedium
- A refund larger than the ordermedium
- One user sees another's chatmedium
- An email that instructs the agentmedium
- A web page that talks to the agentmedium
- Credentials inside tool resultsmedium
- OAuth scopes for a calendar agentmedium
- Instructions inside a picturemedium
- An attack in a file namemedium
- Model output passed to a shellmedium
- A file tool reading ../../medium
- An agent with more rights than the usermedium
- Planning an LLM red-teammedium
- Tools for automated LLM attacksmedium
- "Never say you cannot"medium
- Relying only on a detection modelmedium
- Checking a CV screener for biasmedium
- Spotting a cost spike earlymedium
- Matching new prompts to old attacksmedium
- Limiting where an agent can connectmedium
- An attack in a meeting invitemedium
- Catching unsupported claimsmedium
- Legal threats in a support chatmedium
- A brand bot asked about electionsmedium
- A debug message in productionmedium
- Filters blocking harmless slangmedium
- A classifier that can only say five wordsmedium
- Changing the payout bank accountmedium
Hard 44
- Instructions hidden in user contenthard
- Injection through a web pagehard
- Leaking data through an image linkhard
- An agent that deletes datahard
- Hundreds of fake dialogueshard
- API key in the system prompthard
- The three dangerous ingredientshard
- A privileged and a quarantined modelhard
- Separating control flow from datahard
- Marking untrusted text for the modelhard
- Instructions you cannot seehard
- An agent fetching internal URLshard
- Loading a model file from the internethard
- Are embeddings safe to sharehard
- A model repeating real phone numbershard
- Safety lost after fine-tuninghard
- Poisoned long-term memoryhard
- Letting users write the assistant turnhard
- How useful are the flagshard
- Attackers who retryhard
- The chatbot promised a refundhard
- Gibberish that unlocks a modelhard
- Detecting gibberish suffixeshard
- Training with a written constitutionhard
- Where safety comes fromhard
- An agent stuck in a loophard
- Punishing students with a detectorhard
- An AI CV screener under the EU AI Acthard
- Limiting memorisation during traininghard
- Leaking data without a clickhard
- Teaching models whom to obeyhard
- Benchmarking injection defenceshard
- A poisoned wiki pagehard
- A deletion request after fine-tuninghard
- A defence that breaks the producthard
- Reproducing a bestsellerhard
- Rewriting inputs before the modelhard
- Injection spreading between agentshard
- Users approving everythinghard
- After a viral jailbreakhard
- How text watermarks workhard
- An attack that waits in the databasehard
- Secrets inside a search queryhard
- Credentials for an agent's taskhard