Enforcing Principle of Least Privilege (PoLP) and Separation of Duties (SoD) in day-to-day AI operations rather than just initial architectural design.
2
New cards
AI operational roles (Model Owner vs Data Owner)
Model owners are accountable for a model's operational performance and risk, while data owners are accountable for the data feeding it, establishing clear responsibility paths during incidents.
3
New cards
AI release requirements
Pre-production criteria requiring human override baselines (mandatory human-override capability before release) and reject rate thresholds (ensuring model error and rejection rates stay within acceptable limits).
4
New cards
Operational playbooks, runbooks, and tabletop simulations (AI)
Pre-written incident response procedures for AI-specific scenarios (such as data poisoning and model tampering) paired with regular simulation exercises to validate readiness before a real incident occurs.
5
New cards
Configuration management and drift tracking (AI)
Controlling and tracking changes to AI system configurations and automated model updates while detecting when live configurations silently drift from approved baselines.
6
New cards
AI telemetry data
AI-specific operational data required for security logging—including prompts, completions (model outputs), embeddings, tool calls, and API traffic—expanding traditional network and system logs.
7
New cards
AI event correlation across SIEM, SOAR, and threat intel
Integrating AI-specific telemetry into centralized security tooling so AI events are correlated with broader enterprise security logs rather than remaining siloed.
8
New cards
Log integrity and evidence documentation (AI)
Protecting AI telemetry logs from tampering and preserving them properly to maintain forensic soundness for incident investigations.
9
New cards
Privacy-preserving telemetry masking
Masking or redacting sensitive personal or proprietary information within logged prompts and completions while preserving enough context for effective security monitoring.
10
New cards
Real-time monitoring of AI behavior, drift, and agent actions
Continuously observing live AI systems not only for overt attacks, but also for silent behavioral drift or unexpected autonomous agent actions.
11
New cards
Runtime threat detection (AI)
Inspecting inputs, prompts, retrieval chains, outputs, and resource consumption in real time to detect active attacks or system misuse.
12
New cards
AI access and agency misuse
Runtime detection of unauthorized identities accessing an AI system or legitimate over-privileged agents acting beyond their intended scope and authority.
13
New cards
Secret exposure and tool execution anomalies
Detecting leaked credentials in AI interactions or identifying when an AI agent or connected tool executes in an unexpected or suspicious manner.
14
New cards
AI asset and lifecycle integrity
Continuously verifying at runtime that training data, knowledge bases, models, and deployment pipelines remain unaltered and untampered with.
15
New cards
AI-driven security operations
Using AI and ML defensively in SecOps—such as UEBA (User and Entity Behavior Analytics) for behavioral baselining, automated alert triage, and threat hunting.
16
New cards
Model vulnerability and exposure registers
A tracked inventory of known security weaknesses in specific AI models, mirroring traditional vulnerability management registers.
17
New cards
Context and consequence prioritization (AI vulnerabilities)
Prioritizing which AI vulnerabilities to remediate first based on real-world business impact and operational context rather than relying solely on raw severity scores.
18
New cards
Patch management for AI pipelines and dependencies
Keeping underlying libraries, frameworks, and open-source components within the AI stack updated to remediate known software vulnerabilities.
19
New cards
Compensating controls for legacy models
Applying supplementary security restrictions and monitoring to manage risk when an older or vulnerable AI model cannot be immediately patched or replaced.
20
New cards
Continuous AI security reviews
Ongoing, rather than point-in-time, assessments of AI security posture paired with metrics to verify continuous risk reduction.
21
New cards
AI-specific triage taxonomy
Categorizing AI security incidents by attack type—such as prompt injection, jailbreaking, data poisoning, or model theft—to trigger the appropriate response playbook.
22
New cards
Multi-disciplinary escalation paths (AI IR)
Coordinating incident response across SecOps, MLOps, data science, legal, and compliance teams, reflecting the broader cross-functional impact of AI incidents.
23
New cards
AI-specific containment actions
Immediate measures to stop an active AI incident—such as model rollback (reverting to a known-good version), endpoint isolation, and rate limiting—analogous to isolating an infected host.
24
New cards
Pipeline suspension and agent termination
Halting an active MLOps deployment pipeline or shutting down a misbehaving autonomous agent mid-task to stop ongoing damage.
25
New cards
Preservation of investigative artifacts (AI forensics)
Capturing forensic snapshots of model states, weights, configurations, and datasets—the AI equivalent of preserving a forensic disk image.
26
New cards
Attack kill chain reconstruction (AI)
Tracing how an AI security incident unfolded end-to-end, such as tracking how poisoned data entered a pipeline or how untrusted tool output triggered a compromised action.
27
New cards
RCA, forensics, and timeline tracking (AI)
Applying standard root cause analysis and chronological event reconstruction to AI-specific artifacts and telemetry.
28
New cards
Threat actor attribution (AI incidents)
Identifying who or what caused an AI incident, complicated by the need to distinguish among external attackers, malicious insiders, and manipulated autonomous agents.
29
New cards
Threat eradication (AI)
Removing the root cause of an AI compromise through data sanitization, model retraining, and patching underlying vulnerabilities.
30
New cards
Safety regression testing before reintroduction
Re-testing a remediated AI model prior to production redeployment to verify that the fix works and did not introduce new flaws or break existing functionality.
31
New cards
AI recovery strategies
Restoring normal operations after an incident using trusted model rollbacks, clean checkpoint restorations, and vector-store recovery.
32
New cards
AI circuit breakers and kill switches
Automated or manual mechanisms designed to immediately halt an AI system's actions the moment anomalous or dangerous behavior occurs, limiting real-time damage.
33
New cards
BC/DR plans and critical AI dependency mapping
Mapping which business processes rely on specific AI systems so business continuity and disaster recovery plans account for AI as a critical operational dependency.
34
New cards
Post-incident reviews and TTP register updates
Conducting structured lessons-learned reviews after an AI incident and updating organizational records of adversary Tactics, Techniques, and Procedures specific to AI threats.
35
New cards
Secure AI decommissioning and disposal
Executing machine unlearning and formal model termination to ensure a retired AI model and the sensitive data it learned from are verifiably disposed of rather than merely switched off.