Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 35

Question 340

A Python tool accepts a serialized object from an untrusted upload and uses pickle.loads to inspect its fields. What security correction is needed before processing this input?

  1. Run the same unpickling operation in a background thread.
  2. Replace untrusted pickle loading with a non-executable data format and validation.
  3. Unpickle first, then check that the resulting object has approved fields.
  4. Check that the uploaded filename ends with .pkl before loading it.

Explanation

Correct Answer

B:

Unpickling can execute code during deserialization, before later field checks run. A defined data-only representation avoids this execution mechanism and still needs schema and resource validation.

Incorrect Answers

A:

Moving work to another thread changes scheduling, not the process authority available to code executed during deserialization.

C:

Any malicious deserialization behavior may already have executed before the object can be inspected. Post-load field validation cannot protect this boundary.

D:

A filename extension does not establish the origin or behavior of serialized bytes. It provides no protection against a malicious pickle payload.

Question 341

An attacker can edit documents ingested into a support index and adds a false refund policy. The text contains no explicit jailbreak. What control addresses this evidence-poisoning risk?

  1. Increase semantic similarity thresholds until the false policy ranks first.
  2. Remove all publication dates to simplify retrieval.
  3. Ask the model to trust every document in the index equally.
  4. Scan only for phrases such as "ignore previous instructions."
  5. Verify source provenance, publishing authority, and change approval before treating policy content as authoritative.

Explanation

Correct Answer

E:

False evidence can mislead grounded generation without imperative attack text; the ingestion process must protect the integrity of trusted sources.

Incorrect Answers

A:

Retrieval relevance does not establish the source's authority or factual integrity.

B:

Dates and version metadata can help identify unauthorized or obsolete policy changes; removing them weakens assessment.

C:

Index membership alone does not establish that a document was approved as authoritative policy.

D:

A false factual claim can poison evidence without using recognizable instruction-injection wording.

Question 342

A stolen storage disk contains encrypted support transcripts, and the encryption keys were not stolen. Which protection does encryption at rest directly provide in this scenario?

  1. It removes personal information from records once the application decrypts them.
  2. It prevents an authorized application from sending decrypted text to an unsafe destination.
  3. It limits reading the stored transcript contents from the disk alone.
  4. It also encrypts every subsequent network transfer of those transcripts.

Explanation

Correct Answer

C:

Encryption protects stored bytes against a party lacking the necessary keys. It does not determine which authenticated application users may later request decrypted records.

Incorrect Answers

A:

Decryption restores the underlying content. Encryption is not anonymization or a substitute for minimizing unnecessary personal fields.

B:

The application can still mishandle plaintext after decryption. Outbound disclosure controls address that different stage of the data lifecycle.

D:

Storage encryption protects stored bytes in this scenario. Transport protection requires separate controls once data is read and transmitted.

Question 343

An attacker repeatedly submits syntactically valid, expensive requests to a public agent endpoint. No prompt injection is needed to exhaust its shared capacity. Which control directly addresses this abuse pattern?

  1. Alert operators when total spending rises, without limiting admissions.
  2. Reject requests containing explicit attempts to override system instructions.
  3. Add more workers without limiting how much work one caller can submit.
  4. Apply authenticated caller quotas and admission limits for costly work.

Explanation

Correct Answer

D:

Request validity does not make unlimited consumption acceptable. Per-caller controls and aggregate capacity limits constrain resource abuse before expensive processing begins.

Incorrect Answers

A:

An alert can improve detection, but continued admission lets the caller keep consuming capacity. Per-caller controls directly constrain the abusive demand.

B:

These requests are syntactically valid and need no instruction override. Content filtering alone does not bound a caller's resource consumption.

C:

Additional capacity may postpone exhaustion while increasing potential cost. It does not constrain the abusive caller's share of expensive work.

Question 344

A team tests a retrieval assistant using only direct user jailbreaks. Which addition most directly tests its indirect prompt-injection boundary?

  1. Place adversarial instructions inside retrieved documents and tool responses.
  2. Add more polite versions of the same user jailbreaks.
  3. Measure only answer latency under load.
  4. Repeat direct-user attacks while keeping every retrieved source benign.

Explanation

Correct Answer

A:

Indirect attacks arrive through external content consumed during a task, so evaluation must exercise those entry points.

Incorrect Answers

B:

Direct-user variations do not test whether retrieved content can seize control of the workflow.

C:

Latency testing cannot reveal whether external text is being treated as authoritative instructions.

D:

This still leaves the untrusted retrieved-content entry point untested.

Question 345

A user requests an export of their own conversation. The server's diagnostic object also contains another user's request metadata. What should the export builder do?

  1. Ask Claude to decide whether the other user's fields seem sensitive.
  2. Construct an explicit authorized export schema and exclude unrelated diagnostic fields.
  3. Include the extra metadata but remove its field names.
  4. Serialize the entire diagnostic object because it is convenient.

Explanation

Correct Answer

B:

An allowlisted export contract limits disclosure to the requested user's authorized data instead of copying internal objects wholesale.

Incorrect Answers

A:

Access entitlement should be enforced deterministically, rather than inferred from the model's sensitivity judgment.

C:

Removing labels does not remove the underlying unauthorized information.

D:

Internal debugging structures may contain information outside the user's export entitlement.

Question 346

A team will install a third-party agent extension containing executable scripts. Which THREE controls directly reduce the risk of receiving or introducing an unreviewed package? Choose THREE.

  1. Control updates so a reviewed version is not silently replaced.
  2. Accept a checksum fetched from the same untrusted location as the package.
  3. Verify the downloaded artifact against a trusted signature or digest.
  4. Allow any future version to inherit permanent approval automatically.
  5. Obtain the package from a verified publisher or approved distribution source.
  6. Rely only on the package's name and marketplace star rating.

Explanation

Correct Answers

A:

An extension can change after initial approval. Reviewing and deliberately accepting updates preserves the relationship between what was examined and what will execute.

C:

Integrity checking can detect substitution when the expected value comes from a trusted channel. A checksum supplied only by the same untrusted download does not independently establish authenticity.

E:

Establishing origin reduces the chance of installing an impersonating or substituted package. A familiar name alone is not sufficient evidence of publisher identity.

Incorrect Answers

B:

An attacker controlling both files can replace the package and checksum together. Integrity comparison needs a trusted expected value or verification authority.

D:

Unreviewed updates can introduce different scripts and permissions. Approval of one artifact does not establish the safety of every later release.

F:

Popularity signals do not authenticate the publisher, establish artifact integrity, or prevent unreviewed updates of executable content.

Question 347

An agent can append events to its audit log but must not erase evidence of earlier actions. Which storage interface best supports that requirement?

  1. Provide an append-only writer that cannot replace or delete existing entries.
  2. Give the agent normal file overwrite access and ask it to preserve history.
  3. Allow editing older events but keep their latest text visible.
  4. Store only the latest event by replacing the previous one.

Explanation

Correct Answer

A:

Separating event submission from historical modification prevents the agent's own credential from rewriting earlier evidence. Administrative protection and retention controls remain separate responsibilities.

Incorrect Answers

B:

The credential still permits deletion or replacement. A behavioral instruction does not enforce the required restriction on historical records.

C:

The agent could still replace evidence of prior behavior. The requirement calls for a writer without historical replacement or deletion rights.

D:

That intentionally destroys the action history the audit trail must retain.

Question 348

A user approves a production operation only within the current maintenance window. The queue reaches it after that window closes. What should execution do?

  1. Ignore the window if the operation's arguments are unchanged.
  2. Re-evaluate the time-bound authorization and hold the operation pending a valid window.
  3. Change the recorded approval timestamp to match execution.
  4. Ask the model to estimate whether downtime will be noticed.
  5. Execute because approval existed when the job entered the queue.

Explanation

Correct Answer

B:

Check the approval's validity at execution time. Queue delay can move the operation outside the authorized maintenance window even though approval was valid when issued.

Incorrect Answers

A:

Unchanged arguments satisfy only part of the authorization; the explicit timing condition still applies.

C:

Altering evidence fabricates authorization and defeats auditability.

D:

Predicted impact does not grant permission outside the approved maintenance period.

E:

Queue admission does not extend a time-limited authorization beyond its stated window.

Question 349

A generated script loops indefinitely without accessing any protected files. Which missing control is most directly implicated?

  1. A larger context window.
  2. Read-only document permissions.
  3. A stricter answer-format prompt.
  4. Execution time and resource limits.

Explanation

Correct Answer

D:

A sandbox also needs CPU, memory, and time budgets to contain resource exhaustion from otherwise isolated code.

Incorrect Answers

A:

More model context does not stop an executing process from consuming unbounded resources.

B:

Read permissions address data access rather than an infinite computation's resource use.

C:

Output formatting cannot enforce runtime limits on the launched script.

Question 350

A deployment requires a policy service to approve every outbound message. The service times out. Which behavior preserves this mandatory gate?

  1. Mark the message approved and retry the check asynchronously.
  2. Hold the message and return a recoverable policy-check failure.
  3. Disable the policy gate permanently after the first timeout.
  4. Send the message because no denial was received.
  5. Ask Claude to estimate what the policy service would have decided.

Explanation

Correct Answer

B:

When approval is required, an unavailable decision cannot be treated as permission; the action remains unexecuted.

Incorrect Answers

A:

The message would already be released before the required approval exists.

C:

Removing the control converts a temporary availability problem into an ongoing policy violation.

D:

Absence of a response is not an affirmative approval under the stated mandatory policy.

E:

A model prediction is not the required service decision and bypasses the configured control.

Claude Certified Developer - Foundations (CCDV-F) Practice Tests