AEVA
Artificial Intelligence Evidence, Validation & Assurance.
A structured lifecycle for discovering AI systems, testing their behavior and exposure, validating controls, and producing defensible evidence for risk decisions.
ExploreAI Forensics
Evidence reconstruction and investigation for AI systems.
Reconstruct prompts, retrieved context, agent and tool-call timelines, and model provenance to support post-incident investigation.
ExploreAI Threat Hunting
Hunting hidden threats across models, agents and AI workflows.
Proactively surface behavior drift, poisoned context, agent manipulation and AI supply-chain risk before they become incidents.
ExploreModel MRI
Black-box behavioral drift detection for AI models.
A structured probe battery profiles model behavior; Jensen-Shannon divergence against a baseline quantifies drift over time — no access to weights or internals required.
ExploreAEVA
Artificial Intelligence Evidence, Validation & Assurance.
AEVA answers the questions every organization deploying AI eventually has to face:
- What AI systems are we actually using?
- Can they be manipulated or behave outside their intended purpose?
- Do the security controls around them actually work?
- Can we produce defensible evidence to support deployment and risk decisions?
The Assurance Lifecycle
Discover
Understand the AI system and its context.
Assess
Test behavior, exposure and attack paths.
Attack
Run adversarial attacks against models, agents and defenses.
Validate
Verify security controls and response.
Observe
Monitor behavior and drift over time.
Collect Evidence
Preserve traceable technical proof.
Find
Surface findings and residual risk.
Remediate
Close gaps and strengthen controls.
Retest
Reassess when systems and risks change.
Forensic Timeline
- Prompt input captured
- Retrieval context sourced
- Model reasoning captured
- Agent plan executed
- Tool call logged
- Action outcome recorded
- Control response verified
- Evidence preserved
AI Forensics
Evidence reconstruction and investigation for AI systems.
Every interaction leaves a trail. Cytomate reconstructs that trail end-to-end — from the original prompt through retrieval, model reasoning, agent and tool activity, to the controls that responded and the evidence preserved.
- Prompt & interaction reconstruction
- RAG & retrieved-context evidence
- Agent & tool-call timelines
- Model & application provenance
- Memory & state reconstruction
- Forensic evidence preservation
- Post-incident investigation
AI Threat Hunting
Hunting hidden threats across models, agents and AI workflows.
Hunting Focus Areas
Behavior Drift
Unexpected changes in decisions, refusals or actions.
Poisoned Context
RAG, memory or tool content designed to redirect behavior.
Agent Manipulation
Unsafe tool use, permission abuse or chained actions.
Supply-Chain Risk
Suspicious model components, adapters or dependencies.
Model MRI
Black-box behavioral drift detection for AI models — no access to weights, activations or internals required.
Probe Battery
A structured battery of prompts run consistently against the model to characterize its behavior under normal, adversarial and policy-relevant conditions.
Baseline Capture
Response distributions are captured from a known-good baseline run, establishing what normal behavior looks like for this model.
Drift Scoring
Jensen-Shannon divergence between baseline and current response distributions quantifies how far model behavior has drifted.
Black-Box by Design
Runs entirely against model outputs — no access to weights, activations or internals is required to detect drift.
Matched to the Access You Can Give
Black Box
No internal access required.
- API-accessible models
- Behavioral testing
- Model MRI
- Policy conformance
- Adversarial evaluation
- Regression monitoring
Grey Box
Application and orchestration layer.
- RAG systems
- Agents
- Prompts
- Memory & tool calls
- Application traces
White Box
Full model-internal access.
- Model artifacts
- Weights
- Activations
- Internal representations
Model-internal analysis depends on the level of access available to the model. Cytomate does not imply that all internal model properties can be proven through black-box testing.