What genuinely needs new process
Model and data lifecycle governance, impact assessment, and evaluation of model behavior.
Tap to turn
Operational risk assessment
Oversight, monitoring, incident handling and dependency on the system.
Tap to turn
Operational risk assessment
Oversight, monitoring, incident handling and dependency on the system.
Tap to turn
Excessive agency
A system granted permissions and autonomy beyond its need, so a manipulation becomes an action.
Tap to turn
Secrets handling in prompts and context
Never place credentials in prompts, since context can be revealed by output.
Tap to turn
Culture signal in AI governance
Whether teams raise concerns before launch or after the first complaint.
Tap to turn
Risk categories specific to AI
Data, model, operational, third-party, regulatory, ethical and reputational.
Tap to turn
Risk categories specific to AI
Data, model, operational, third-party, regulatory, ethical and reputational.
Tap to turn
Prompt injection
Malicious instructions in content the model processes, overriding its intended constraints.
Tap to turn
Communicating an AI incident
Facts confirmed, affected parties identified, and regulatory obligations assessed.
Tap to turn
Communicating AI risk to executives
In business consequence, with the decision required stated plainly.
Tap to turn
Why generic AI risk statements fail
"AI risk" cannot be owned, assessed or treated.
Tap to turn
Why generic AI risk statements fail
"AI risk" cannot be owned, assessed or treated.
Tap to turn
Why retrieval augmentation matters to security
It moves the control point to the retrieval corpus and its access permissions.
Tap to turn
Containment options for AI systems
Disable the feature, restrict the tool permissions, tighten guardrails, or roll back the version.
Tap to turn
Previous 1 of 15 Next card
Question 1 of 10 · AI Governance and Program Management
Which data should staff be told not to enter into public AI tools?
A Confidential, personal or regulated data B Public product descriptions from the company website C Generic writing requests D Questions about public topics
Public tools may retain or train on inputs, so confidential and personal data should not be entered.
Next question
Question 2 of 10 · AI Risk Management
What should be done when an AI vendor cannot provide evidence of security testing?
A Treat it as a risk: compensate or reconsider the vendor B Accept the vendor's verbal assurance and proceed C Assume the testing was done and record no finding D Run a scan against the vendor's public website
Missing assurance evidence is itself a risk finding: the organization should assess it, apply compensating controls such as contractual commitments or its own testing, or choose another vendor. Verbal assurance and assumptions are not evidence, and scanning the vendor's website says nothing about the AI system's security.
Next question
Question 3 of 10 · AI Technologies and Controls
What should be tested before deploying an updated model version?
A Only that the new model file loads without errors B Performance, safety and security against prior baselines C Only the response speed and latency of the new version D Whatever the vendor's own release notes recommend
An updated model can regress in accuracy, safety behavior or resistance to attacks such as prompt injection, so it should be regression tested against the previous version's baselines before deployment. Checking that it loads or measuring speed alone misses those regressions, and vendor release notes do not reflect the organization's own use and controls.
Next question
Question 4 of 10 · AI Governance and Program Management
What should AI policies say about third-party AI features embedded in existing software?
A They are subject to assessment like any other AI use B They are exempt because the software was already approved for use C They are the vendor's problem only D They must be disabled
New AI features in existing tools change data flows and risk, so they need assessment.
Next question
Question 5 of 10 · AI Risk Management
What is a key risk in using AI for security operations?
A The AI will take too long to read the organization's log files B Higher storage costs C Over-reliance on outputs without analyst verification D Fewer alerts
Automation bias leads analysts to accept wrong conclusions; verification remains necessary.
Next question
Question 6 of 10 · AI Technologies and Controls
Which attack corrupts training data to change a model's behavior?
A Data poisoning B Model extraction, which copies a model by querying it C Membership inference D Evasion
Poisoning alters training data to degrade or backdoor a model; extraction and membership inference target a trained model.
Next question
Question 7 of 10 · AI Governance and Program Management
What should an AI governance charter define?
A The committee's scope, authority, membership and decision rights B The detailed architecture of each AI system the organization runs C The list of approved prompts D The model training schedule
A charter sets mandate, scope, authority and decision rights; technical detail belongs elsewhere.
Next question
Question 8 of 10 · AI Risk Management
How should AI risks be recorded?
A In a separate spreadsheet kept only by the data science team B In the model code C Not at all D In the enterprise risk register with owners and treatments
AI risks belong in the enterprise register for visibility and accountability.
Next question
Question 9 of 10 · AI Technologies and Controls
Which technique reduces the risk of a model memorizing personal data?
A Minimizing and de-identifying training data, or differential privacy in training B Training for more epochs so the model generalizes better from the data C Using more data D Increasing model size
Minimization, de-identification and differentially private training reduce memorization.
Next question
Question 10 of 10 · AI Governance and Program Management
What should a security manager do FIRST when asked to secure the organization's use of AI?
A Build an inventory of AI systems and how they are used B Deploy an AI firewall in front of every model the organization uses C Ban all generative AI tools until a policy is written D Buy a red-teaming service
Security cannot be applied to systems nobody knows about. An inventory of models, tools and uses, including shadow AI, comes first.
See my score
0 of 10
AI Governance and Program Management
AI Risk Management
AI Technologies and Controls