Viskor finds, explains and controls every use of AI on your network: people chatting with AI, programs calling AI APIs and autonomous agents using tools. It shows what each one touches, who is behind it, and what changed. No endpoint agents required, and no changes to your stack.
Developers, analysts and power users all run AI: Copilots, custom LLM workflows, autonomous coding agents, MCP tool servers. Most security teams cannot see what is calling out, to whom, or on whose behalf.
SIEM and firewall rules weren't written for LLM endpoints. Agent calls blend in with ordinary web traffic.
Agents send code, documents and emails to outside models, and prompt injection can make them do it on an attacker's behalf.
Raw IP logs don't map back to people, teams or business context, and agents have no owner at all.
Approved tools change weekly, blanket bans push AI underground, and regulators now ask for an inventory and evidence.
DNS, proxy, browser extension and your existing logs and traces each see a different part of AI traffic.
Every event gets a provider, a confidence that it is AI, and a separate confidence that it is an agent at work.
DHCP and LDAP / Active Directory turn addresses into a person, department and computer.
Policies allow, alert or block, with alerts delivered to your SIEM, chat or ticketing tools.
Behaviour detection compares each person, agent and department with its own history and raises Findings.
Everything a security team needs to discover, understand, control and evidence AI use, in one portal.
Shows which AI services every device looks up. Covers every app with no content inspection, and can block lookups.
Sees the requests themselves: model, tokens, tool calls and agent names. Enforces block decisions and rate limits.
Records AI web chats in ChatGPT, Claude, Gemini, Copilot and Perplexity. It sends usage details only, never the text of prompts or replies.
Reuse the resolver, firewall and gateway logs (syslog or HTTPS) and OpenTelemetry traces you already collect. No new sensors needed.
Add any product with a JSON format definition and a few sample lines, tested before you save.
Declare self-hosted models and gateways (Ollama, vLLM, your own) so their traffic is recognised by name, even when encrypted.
Every AI event with full detail, and a plain-language explanation of why it counts as AI or agent activity.
Events grouped into conversations and agent runs, with a timeline of every model and tool call in order.
DHCP and LDAP / AD sources and an IP user map put a person, department and computer behind every address.
Allow, alert or block by provider, model, MCP tool, subnet, confidence or risk, with a custom block message and token rate limits.
Signed webhooks send policy alerts and behaviour findings to your SIEM, chat or ticketing system, filtered by risk and severity.
Advanced Agent Detection, internal (east-west) traffic observation and optional internal TLS decryption, each switched on independently.
Claude Code, Codex, OpenAI SDKs, browsers and more, identified from the software. Agents can also name themselves with a header.
See which Model Context Protocol tools an agent called, in order, inside each session.
Two confidence scores: is this AI, and is it an agent at work (tool use, MCP, machine-speed loops)?
Every event records what the sensor could observe (connection, headers or body), so gaps are explicit.
Every AI client and agent, with its providers and activity. Duplicate identities are suggested for merging.
Assign an accountable owner and a friendly name. Network identity suggests a likely owner.
See the models an agent uses, the tools it calls (grouped by what they do) and the agents and servers it reaches, including a second hop.
A live map and Sankey view of agents, providers, MCP servers and internal hosts, including agent-to-agent links with confidence scores.
A wide set of detectors compare each person, department, agent and computer with its own history after a learning period. Acknowledge, dismiss or suppress a finding, or put someone on a watchlist so stricter policies apply.
A person or team uses an AI service for the first time.
An agent identity never seen before.
An agent switches to a model it has never used.
An agent picks up a tool or server it has never used.
Active again after 30 or more days of silence.
A computer starts talking to an AI service nobody recognises.
Activity at an hour or on a day they never use.
An owned agent run by a different person.
At least three times a typical day's tokens.
Far more data sent to AI than usual.
Everyone together, even when no one person stands out.
Usage doubled compared with a month ago.
A session ten times its usual size: a loop that won't stop.
Reads private data or runs code, then sends something out within 30 minutes.
An outside model gets to a sensitive internal server through another agent.
An agent talks to another agent or server for the first time.
Many new internal destinations in a day: the shape of lateral movement.
Several agents together reaching places nobody reached before.
Reached the provider another way, or switched provider, within 15 minutes of a block.
A person's account starts behaving like an agent.
An activity map of departments and their agents, with stat cards, trends and breakdowns. Click to zoom in, or press Present for a big screen.
Four headline numbers and the activity map, with no tables and no jargon. Made for senior audiences.
An action queue: high-risk findings, new agents, agents with no owner, quiet data sources and recent blocks, with one-click actions.
Everything Viskor knows as an interactive 3D globe of people, agents, models, tools and findings, joined by their real relationships.
A compliance module aligned with the EU AI Act turns observed AI use into the inventory, risk view and decision record that regulators and auditors ask for.
Which AI providers each department uses, with users, events and risk tier. Rebuilt nightly, and exportable.
Your own risk tier, GPAI flag and regulatory tags for each provider, with a written rationale and full history.
Viskor flags possible incidents but never escalates on its own. A person decides, with a note, and every decision is logged.
CSV evidence for any period, and a record of who changed what in the portal.
Viskor serves organisations worldwide. Run it on any cloud infrastructure or in your own data centre, so you choose where your data lives. You get deep visibility without friction: nothing to install on endpoints and nothing to re-architect. Every operator has their own account, and every change is audited.
Interactive walk-throughs of the portal. They run in your browser on illustrative example data.
Thirteen steps through the portal, from dashboard to agent graph.
Spot an unapproved agent and stop it with a policy.
See an agent change behaviour and what Viskor reports.
Follow the chain from untrusted input to sensitive data.
Replay what happened, event by event.
Viskor gives every AI agent a stable identity, a named owner and an audit trail. Shadow AI becomes something you can govern, not just watch. We're onboarding a small group of security and identity teams for early access.
We use your email only to contact you about Viskor early access. It is never shared, and we don't send spam. Ask us to delete your details at any time: hello@viskor.ai.
