AI Security and Safety; Cloud and AI Security
Free CSA Trusted AI Safety Expert (TAISE) practice — 6 questions on AI Security and Safety; Cloud and AI Security, with explanations. No sign-up. Full 12-question mixed test →
A cloud provider hosts an AI inference platform where multiple tenants upload and run their own fine-tuned LLMs on shared GPU clusters. During a security review, you must select a control that prevents one tenant's inference workload from reading another tenant's model weights or KV-cache memory residue. Which control BEST addresses this risk?
Cross-tenant memory and weight leakage (including GPU memory residue and side-channel attacks) requires hardware/kernel-level isolation, not just logical access control. Dedicated sandboxed runtimes with memory scrubbing prevent data remanence and side-channel exposure, which is the TAISE-tested control for securing multi-tenant AI workloads on shared compute.
An agentic AI assistant browses the web and can call a 'send_email' tool. An attacker plants hidden instructions in a webpage the agent retrieves (indirect prompt injection), causing the agent to exfiltrate conversation history via the email tool. Which mitigation strategy BEST reduces this risk?
Indirect prompt injection exploits the agent treating retrieved content as trusted instructions. TAISE's defense-in-depth approach combines least-privilege tool scoping, human-in-the-loop for high-impact actions, and strict instruction/data separation with output validation — this layered control set is the tested mitigation for agent/tool security.
During red-team evaluation of a production LLM API, you suspect an adversary is attempting a model extraction (model stealing) attack by systematically querying the endpoint. Which detection approach is MOST effective for identifying this activity?
Model extraction attacks require large volumes of structured probing queries to reconstruct decision boundaries or weights. TAISE-tested detection relies on behavioral analytics (query pattern/anomaly detection) paired with rate limiting and per-key quotas to disrupt the attack before sufficient data is exfiltrated.
Your organization consumes a fully managed, provider-hosted foundation model API (no self-hosted infrastructure). Under the cloud shared-responsibility model for AI services, which of the following remains the CUSTOMER's responsibility?
In the shared-responsibility model for fully managed AI services, the provider owns infrastructure, host/OS patching, physical security, and model weight protection, while the customer remains responsible for credential management, input/output handling in their application layer, and configuring IAM/access policies around their usage — a frequently tested boundary on TAISE.
Which statement BEST distinguishes AI red-teaming from AI safety benchmark evaluation?
TAISE distinguishes red-teaming (creative, adversarial, often human-driven probing for unknown vulnerabilities and novel attack paths) from safety benchmarking (automated, standardized, reproducible scoring against fixed datasets/metrics for known risk categories like bias, toxicity, or jailbreak resistance). Both are complementary components of a safety evaluation program.
A secure MLOps pipeline pulls a third-party open-source training dataset and a pretrained base model from a public model hub. Weeks after deployment, you discover the model exhibits a backdoor trigger that was injected via a poisoned dependency during the build process. Which control would have BEST prevented this supply-chain compromise?
Supply-chain attacks on training pipelines (poisoned dependencies, tampered datasets, or backdoored base models) are mitigated through provenance attestation and signing at every stage (data, code, dependencies, artifacts), combined with reproducible build verification — this is the TAISE-tested secure MLOps control to prevent undetected injection before the model reaches production.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →