Runtime control plane of the quality management system
The AI Analyzer is the runtime control plane of the quality management system. It receives data from gateways and agent harnesses and evaluates it for trustworthy AI.
Ingests prompts, responses, hashes, trajectory identifiers, turn indexes, optional runtime snapshots, and system or agent identifiers. Events, metrics and an ingestion map can be listed. Retention can keep hashes only and discard original text.
Applies rules for jailbreaks, indirect prompt injection, content patterns, prompt and orchestration drift, behavioural indicators, shadow agents, network and SaaS shadow AI, honey tokens and OWASP AIVSS scoring.
Before a tool runs, authorisation checks enforce MCP allowlists, high-risk tool combinations, irreversible-action guards, financial limits and intent checks across phases. In fail-closed mode a critical violation halts the harness. Other outcomes are log-only or escalate.
Provides a human-oversight queue with justification, human-in-the-loop service levels, signed evidence packs, a hash-chained governance audit log, bias and explainability companions, and watermark detection.
Generates control-mapping and conformance exports (including ISO/IEC 42001), compiles agent configurations, produces software bills of materials and lockfiles, and synchronises the inventory. Alerts can be routed to the governance, inventory, oversight, watermark and monitoring modules.
It is not a SIEM, not a model-internal guardrail and not a chain-of-custody forensics store.
