What is AI security, and why does it need a dedicated platform?
AI introduces an attack surface traditional tools miss — prompt injection, jailbreaks, data leakage through prompts, autonomous-agent misuse and model-supply-chain risk. A dedicated platform inspects AI traffic, ties it to identity, and maps it to AI-specific threat frameworks.
Does Kapālins cover the OWASP LLM Top 10?
Yes. Guardrail controls and model probes map to the OWASP LLM Top 10 — prompt injection (LLM01), insecure output handling, sensitive-information disclosure, excessive agency and more — and every finding is cross-walked to OWASP alongside MITRE ATLAS and NIST AI RMF.
Does Kapālins support the NIST AI Risk Management Framework?
Yes. Kapālins assesses and scores your posture against the NIST AI RMF functions (Govern, Map, Measure, Manage), recommending and scoring controls from your policies and live telemetry — so the framework becomes a running, evidence-backed score, not a one-off questionnaire.
Does Kapālins map to MITRE ATLAS?
Yes. AI threats and detections are mapped to MITRE ATLAS, the adversarial-ML knowledge base, so you can see which adversary techniques apply to your AI systems — viewed in the Command Center alongside OWASP and NIST.
Does Kapālins do AI red teaming?
Yes. AI red-team and adversarial testing probe your models and agents for jailbreaks, injection and unsafe behaviour, with findings fed back into your posture and evidence chain.
How does Kapālins prevent data leakage (DLP for AI)?
Data Shield inspects prompts and responses for PII, secrets and cross-tenant data. In Detect mode it flags exposure; in Enforce mode it redacts or blocks before sensitive data reaches the model or leaves it.
Does it cover model integrity and supply chain (Model SBOM)?
Yes — a Model SBOM (software bill of materials for AI) tracks the models, versions and dependencies in use so you can manage model-supply-chain risk. It is part of the Fortress tier.
Can Kapālins discover shadow and untested AI across our estate?
Yes. The Probe Runner probes your model endpoints and the AI systems your teams stand up on a schedule, surfacing shadow and untested AI and scoring each (approve, conditional or block) — so it shows up on the Trust Index instead of in an incident report.
Can we run in detect-only mode before enforcing?
Yes. Every shield runs in Detect (observe) mode first, so you can see real risk on live traffic before switching any control to Enforce (block).
How do AI security findings become audit evidence?
Every control maps to compliance frameworks, and posture uses a dual-confidence model — attested from assessments and verified from live gateway telemetry — so findings turn directly into regulator-ready evidence.