AI is no longer an experiment — it's training pipelines, agent workloads, vector stores and MCP servers running with privileged identities. OnDuty AISPM continuously discovers your full AI estate, maps it into one live graph, and surfaces the toxic combinations attackers actually exploit.
Six capabilities, one platform
Inventory every model, agent, dataset, prompt template and MCP server — managed, self-hosted or shadow. Refreshed continuously from runtime, not a one-off scan.
Detect indirect prompt injection, jailbreak attempts and data exfiltration in agent traffic. Map back to the exact workload, identity and tool that was abused.
Find PII, secrets and regulated data inside training sets, embeddings and RAG sources. Flag poisoned datasets and over-permissioned vector stores.
Agents act with IAM. We graph every model, agent and MCP server to its identity, secrets and reachable resources — and score the toxic combinations.
MCP servers are the new attack surface. Audit every server, signed tool, and permitted action — and stop agents from invoking dangerous combinations.
Validate the integrity of model weights, fine-tuning jobs and inference containers — from training notebook to production endpoint.
How the platform works
Stop stitching together a model registry, an agent framework dashboard and a separate vector store audit. OnDuty puts every model, agent, dataset and MCP server on a single data model — so a policy written once applies everywhere, and shadow AI has nowhere to hide.
Unified data model
One live AI graph
Every finding is scored against the agent's identity, the data it can reach and the tools it can invoke — not severity alone. OnDuty surfaces the handful of risks where an over-privileged agent sits next to sensitive data or a dangerous tool, so your team stops triaging alerts that lead nowhere.
Ranked by dynamic risk score
toxic combinationagent://support-copilot reaches prod PII store
Top of 214 open findings
OnDuty automatically clusters related alerts into a single remediation Mission with a clear owner and deadline. One over-permissioned IAM role granted to a shared agent identity can trigger hundreds of findings — OnDuty groups them so you fix the role once, not the symptom a hundred times.
Root cause: shared IAM role agent-runtime-role grants AdministratorAccess to every LangGraph agent that assumes it.
1 role
Fix once
86 alerts
Resolves
12
Agents scoped
OnDuty connects the dots between the notebook that trained a model, the registry it shipped through and the agent it's running in today. A live detection traces straight back to the fine-tuning job or training set that introduced it, so your ML team fixes the source, not just the symptom.
train.ipynb
Fine-tune job · PII in dataset
Model registry
Blocked · sensitive data tag
billing-bot
Deployed agent
Runtime finding
PII reachable via tool call
OnDuty AI shortens the distance between finding a risk and resolving it. Ask a question and get back a query, not a support ticket. Chat with your entire AI estate. And let AI agents handle the analysis and legwork, so your team spends its time on judgment calls instead of repetitive investigation.
AI Policy Fixes
Generated patches for IAM & tool policies
AI Discovery
Ask questions, get asset queries
AI Assistant
Chat with your AI estate
AI Agents
Investigate and act autonomously
The AISPM lifecycle
Agentless scan of cloud, K8s and CI to inventory every model, agent and dataset.
Score each AI asset against OWASP LLM Top 10, NIST AI RMF and EU AI Act controls.
Continuously watch agent traffic and MCP calls for injection, exfiltration and abuse.
Prioritized findings ship with runtime evidence, owner and a one-click remediation path.
FAQ
AISPM (AI Security Posture Management) is a security category focused on discovering and governing the full AI estate — models, autonomous agents, training data, vector stores and MCP servers — rather than treating AI as just another workload. It combines inventory, identity mapping and runtime detection into one graph purpose-built for how agentic AI actually behaves.
Related solutions
OnDuty is one platform on a single graph. Here's where the rest of it connects.
Replace shadow AI with one graph
See your real AI inventory, identity graph and top toxic combinations in a 30-min guided demo with our team.