Free CY0-001 practice test: 30 questions with answers

Prompt injection manipulates the model through input text to override instructions or leak hidden context. The other attacks target databases, browsers or network capacity rather than the model's instructions.

Indirect prompt injection hides instructions in content the model retrieves, so the model follows them without a direct attacker prompt. Model extraction, image patches and credential stuffing are different attack types.

Tool calls should run with the user's permissions and narrowly scoped credentials, so the model cannot exceed them even if manipulated. A shared admin key and unchecked API choice give the model excess power, and removing authentication removes the identity check.

Model extraction uses query responses to reconstruct or imitate a model's behaviour. Backdoor insertion occurs in training, evasion uses a crafted input against a deployed classifier, and session fixation targets web sessions.

Poisoned training samples that implant a hidden trigger create a backdoor in the model, which activates only when the trigger appears. Denial of wallet is cost abuse, and clickjacking and DNS spoofing are web and network attacks.

Output filters and response constraints prevent disclosure of sensitive internal information, which is a common AI security control. Longer outputs increase exposure, and disabling logs or validation makes detection and defence weaker.

Pinning a known model version and re-testing before any upgrade keeps behaviour and security posture predictable. Accepting updates blindly, dropping tests and skipping staging all remove the control that would catch regressions.

A secrets manager with runtime injection keeps keys out of code and prompts, and scanning for credential patterns catches leaks before release. Placing keys in prompts exposes them to the model's outputs, public commits leak them, and shared keys widen the blast radius.

Correlating related alerts and scoring their risk reduces noise and helps analysts focus on the likely incidents. The other capabilities are unrelated to SOC triage.

Generated summaries can omit or invent details, so a human should verify them against the source logs before regulatory use. Publishing without review, truncating the input and raising temperature all reduce reliability.

Anomaly detection models learn a baseline of normal activity and flag unusual patterns, which can reveal unknown attacks. Signature lists, yearly reviews and subject filters only match known patterns.

Testing against labelled events that include benign activity measures both false positives and missed detections, which is the reliable check. Production testing risks alert floods, word counts say nothing about accuracy, and asking the model to self-assess is unreliable.

Out-of-band callback verification and dual approval stop a convincing voice or email from authorizing payments on its own. Trusting voice alone, disabling filters and publishing voice samples all increase the risk.

The useful measure is faster triage of real issues without missing critical ones. Scan counts, summary length and parameter counts do not show whether triage works.

A second analyst catches logic errors that the generator introduced, and testing against labelled benign and malicious events exposes false positives and misses before production. Immediate deployment, disabling logs and suppressing alerts all hide problems rather than solving them.

Retrieval restricted to approved playbooks, with citations and refusals when nothing matches, keeps answers grounded and auditable. General internet models and memory-only answers are not tied to approved content, and generating new playbooks on demand bypasses approval.

A risk assessment that covers impact on customers, bias testing and human oversight is the standard pre-deployment governance step. Publishing weights, skipping review and waiting for a regulator leave the risks unmanaged.

Patient data requires contractual and privacy controls such as a data processing agreement before it is shared with a vendor. Sending test samples, relying on marketing and deletion after a breach all skip this requirement.

Allowing challenge and human review, supported by transparency about how decisions are made, is a core governance principle for automated decisions. Eliminating human involvement, keeping decisions in model memory and hiding logic all work against it.

The NIST AI RMF organizes AI risk work into govern, map, measure and manage functions. PCI DSS covers payment card data, ISO 9001 covers quality management, and HIPAA's breach rule alone does not structure AI risk management.

Human review, software composition analysis for licences and vulnerabilities, and a clear policy give the organization control over AI-generated code. Accepting code unread, dropping licence checks and hiding AI use all increase risk.

Data loss prevention can detect sensitive content leaving the organization and block or warn on unapproved AI destinations, which enforces the policy. A memo alone is not enforcement, blanket blocking harms productivity, and asking users to delete data after the fact does not prevent the exposure.

A disparity in error rates is a fairness risk that requires investigation, documentation and mitigation with oversight. High overall accuracy can hide harm, and removing reports or publishing without a plan does not reduce the risk.

Supervised learning uses labeled examples such as known fraud and legitimate transactions, while unsupervised methods such as clustering look for patterns without labels. The distractor about unsupervised labeling reverses the definition, and supervised learning is widely used for classification.

Grounding answers in retrieved passages with citations and refusing when no supporting source exists reduces hallucinated policies. Raising temperature increases randomness, and removing the index or broadening training does not ground the answers.

With heavy class imbalance, accuracy can look high while missing most malware, so precision, recall and F1 show detection quality for the minority class. Epoch count and file size do not measure detection performance.

A model card records intended use, data sources, limitations and evaluation results, which helps reviewers judge fit for purpose. It does not replace testing, does not hold keys, and does not configure firewalls.

A large gap between training and unseen-data performance suggests the model has memorized patterns in the training set rather than generalizing. Underfitting shows poor results on both sets, and firewall or hardware issues are unrelated to model accuracy.

Embeddings map content into numeric vectors so similarity search can find related items, such as similar alerts or documents. Signatures, firewall rules and permission lists are not embeddings.

Removing direct identifiers and generalizing quasi-identifiers such as dates or postcodes lowers re-identification risk. Copying the dataset widens exposure, while renaming or reformatting the file changes nothing about the content.