+ November 2026
- ImprovedDefense
Protection trends, week on week
Each protection shows what it caught per day and how this week compares with the week before, and the headline total does too.
- NewDefense
Approval health: which rules are worth a person's time
For every rule that holds calls for a person: how its holds ended, how fast people decided, and what to change. Rules approved nearly every time, or left unanswered, are called out with advice.
- ImprovedThreat intel
Threat intel loads fast at any range
Counts are computed in the database, so the 30-day view loads about as fast as the last 24 hours. Filter by technique, agent, outcome, session or incident.
- ImprovedAgents
Review a new agent in one place
The learning review is on the agent's overview, not folded away: what the agent was seen using, how close it is to ready, and the rules to lock in.
- ImprovedPlatform
Seamless releases
A page you had open keeps working through a release: files from the version you loaded are still served, so nothing breaks mid-update.
- NewThreat intel
Threat intel rebuilt around incidents and techniques
Detections are grouped into incidents and mapped to OWASP LLM, OWASP Agentic and MITRE ATLAS techniques. See what your policy does about each technique and block it in one click, recorded in the change log.
- ImprovedDefense
Defense shows what protects your agents
See how many agents are enforced or only watched, what each protection covers and your OWASP LLM Top 10 coverage, with a link to the setting that closes each gap.
- NewConnect
Connect any agent with one MCP URL
Each agent gets one MCP URL, its runtime key alone configures the SDKs, and a new agent's first calls are no longer stopped by setup holds or a key every tool call would refuse.
- NewCLI
Create and onboard agents from itr
itr agent create sets up an agent with its model and a dollar budget, and itr agent onboard connects it. The Connect screen uses the same path. Installers now ship with release checksums.
- NewDiscovery
Govern a discovered fleet in one command
From Discover, one itr command points discovered agents at Intertrace without code changes, and provider keys stay in your vault. Each agent's map shows where it runs and what its cloud permissions reach.
- ImprovedApprovals
Approved calls resume the agent's run
A held call returns a structured hold, and the SDK resumes the run once a person approves, with adapters for Vercel AI, OpenAI Agents and LangGraph. The same call is never held twice.
- NewAgents
See everything an agent holds and touches
Each agent's map and connection graph show its capabilities, how Intertrace knows them and what each tool actually touched. Import a manifest or agent-builder export, and spot drift between discovery scans.
- NewPolicies
Flag a tool: it runs, then gets reviewed
Alongside Allow, Ask and Block, a tool on an agent's Connections can now be set to Flag. The call runs and is marked for review, either recorded or opened as a Work finding.
- NewSDK
LangChain v1 middleware and Google ADK for TypeScript
New LangChain v1 middleware for TypeScript and Python, Google ADK support for TypeScript, and pre-tool hooks that check every tool call. In Python, a block now reads as a plain sentence.
- NewCLI
Put local stdio MCP servers behind Intertrace
itr mcp bridge starts a local stdio MCP server and relays its traffic, so every tool call, resource read and prompt is decided and recorded before it reaches the server.
- SecurityRuntime
Key rotation policy
Set a maximum age for each type of key, get reminders when one is due, and rotate a runtime key in one click with a grace period for the old one. The audit log keeps which key was revoked.
- NewApprovals
Approve with changes and message the agent
Approvers can change a held call's arguments, such as the amount or recipient, and approve only the edited call. They can send the agent a message with their decision and get phone and desktop notifications.
- NewMCP
MCP sign-in with your own OAuth app
For MCP servers that only accept pre-registered apps, use your organization's own OAuth app or a Client ID Metadata Document. Each person can also use their own account behind a server.
- NewMCP
30 ready-made connectors and a bigger MCP catalog
Ready-made connectors for Google Workspace, Microsoft 365, Salesforce, Zendesk, ServiceNow and 25 more. The MCP catalog adds 46 official remote servers, each checked live, and can now be searched.
- NewPlatform
Custom roles, SCIM groups and scoped visibility
Provision users and groups with SCIM, narrow a built-in role into a custom one, use groups in policy conditions and as approvers, and limit what members see to their team's or their own agents.
- SecurityRuntime
Automatic stop, rate limits and call caps
Set rules that stop a rogue agent on its own (OWASP Agentic ASI10), limit requests a minute per runtime key and person, and cap calls to a tool or server per conversation, hour or day.
- NewSimulation Lab
Run Labs in CI
itr labs run tests your agents from CI and sends findings to code scanning as SARIF. Labs can also target any HTTP endpoint from a cURL command, with custom packs and verdict overrides.
- SecurityPlatform
Customer-managed keys
Your organization's own key-encryption key can wrap its secrets in Intertrace. Prompt storage has its own switch under Settings → Security, and stored conversations are encrypted.
- SecurityPlatform
Tamper-evident audit log and signed receipts
The audit log is hash-chained with signed checkpoints, and decisions carry signed receipts. Verify them with itr and download receipts from Settings and Reports.
- NewIntegrations
SIEM export in OCSF, plus OTLP
Send decisions as OCSF 1.8.0 to Splunk, Datadog or Microsoft Sentinel, export hourly to S3, or stream over OTLP.
- NewCosts
Budgets per key, person and team
Set budgets and token caps per runtime key, person and team, with hard caps that fail closed. Add your own model prices and label requests, and exact cache hits are never billed.
- NewDiscovery
Shadow AI from browsers, DNS and compliance logs
The browser extension can check what people type, paste or attach on AI apps before it's sent (opt-in). Discovery also reads DNS logs, GitLab, Bedrock invocation logs and enterprise AI compliance APIs.
- SecurityAgents
Developer machines and what coding tools load
Machines lists the developer machines running itr and flags risky coding-agent modes. Skills, plugins and MCP servers are inventoried and hash-pinned, and packages with a known malicious release are refused.
- SecurityRuntime
Computer-use and browser actions governed
Clicks, typing and navigation from computer-use and browser tools are read from their arguments, so policy decides on what the action does, not on a generic tool name.
- NewReports
Compliance mappings and Agent Assurance Report
Map controls to OWASP Agentic Top 10 2026, OWASP NHI, ISO 42001, HIPAA, PCI DSS, GLBA, NYDFS, CCPA and AIUC-1. A new per-agent report bundles its evidence with verified decision receipts.
- NewIntegrations
Screening API for the gateway you already run
Your own gateway can ask Intertrace about a prompt, reply or tool output, with adapters for LiteLLM, agentgateway, Kong, Portkey, Cloudflare Workers, Envoy, Azure APIM and Apigee.
- NewMCP
MCP resources, prompts and changed tools governed
MCP resources and prompts are governed like tools, a changed tool definition asks a person and shows the diff, and approvals can be answered inside the MCP client.
- NewThreat intel
Intertrace Threat Wire
Briefings on threats to AI agents, written by Intertrace, with recommended actions first and a prompt to check your exposure. Available to every organization, with or without traffic.
- SecurityRuntime
Hidden instructions and prompt leaks stopped
Hidden instructions in tool output are cut out before the model sees them, and tool-output checks no longer fail open when busy. Replies that reproduce the system prompt are withheld.
- NewPolicies
Policy as code
Manage policy in a file: preview it, test it, apply it in one transaction and gate it in CI. Rules such as two-person approval can't be weakened by one person alone.
- NewIntegrations
Copilot Studio, Agentforce and ServiceNow (beta)
Discover and protect agents built in Microsoft Copilot Studio, Salesforce Agentforce and ServiceNow. Your organization's audit log can now be exported.
- NewMCP
Turn any HTTP API into governed MCP tools
Import an OpenAPI or Swagger spec and the gateway serves it as MCP tools, with OAuth to the API and a test console. Bundle tools across servers and manage API tools as code with itr.
+ October 2026
- NewMCP
Hosted MCP servers
Run MCP servers on Intertrace from the dashboard and reach them through the gateway. Each hosted server can only reach the hosts it declares.
- NewPolicies
Organization rules in plain English
Write the rules every agent must follow, one sentence each. See the rule each becomes and what it would have done over the last 30 days before you apply it.
- NewAgents
Describe an agent, get its policy
Describe what an agent does and the builder drafts a policy for it. Edit any line, see sample calls decided under it, and approve. Money, deletes, deploys and outbound messages never run on their own.
- NewPlatform
HIRO, the AI security analyst
Ask HIRO about your agents, sessions and findings and get answers grounded in your records. HIRO can preview and make policy changes, each checked, logged and undoable.
- NewSimulation Lab
Red team for MCP, A2A and agent actions
Labs red-teams MCP servers, A2A agents and agent actions in the background, scans MCP tools for poisoning, re-tests on a schedule and gives each agent one exposure score.
- NewPolicies
Learning period for new agents
New agents start in a learning period: Intertrace records their behaviour, proposes a policy with measured readiness, and locks it in when you're ready. Consequential tools still ask a person.
- NewDiscovery
Discover the AI running in your estate
Discovery scans AWS, GitHub, identity sources and AI vendor orgs, maps what each AI system can reach and shows what changed since the last scan. Optional code read covers Lambda, server disks and containers.
- NewPolicies
Roles that decide what asks a person
Pick a role for each agent and it decides what needs approval by consequence. Work freely in development, add expiring exceptions, require two approvers and pre-approve routine calls.
+ September 2026
- NewPolicies
A tool grant for every agent
Each agent now gets its own compiled tool grant on Policies, instead of one plan for the whole organization. Unknown tools are no longer suggested for allow — they stay denied until someone decides.
- NewWebsite
Videos, explainer and motion across the site
A new video library with product walkthroughs, threat primers, developer quickstarts and industry cuts, plus a 3½-minute explainer on the home page.
- NewSimulation Lab
Test policy packs in the Simulation Lab
Pick a policy pack and run attacks against it before you apply it to real agents. Enforcement in the Lab now matches production.
- NewMCP
MCP Guard
A dedicated overview of how Intertrace authorizes every MCP tool call, inspects tool descriptions, pins schemas and fails closed.
- NewInsights
Technical papers, threat news and guides
The blog now covers runtime AI security in depth, with architecture diagrams, cited sources and an RSS feed.
- SecurityRuntime
Runtime keys bound to environments
Keys are tied to the environment they were created for, and every action keeps its declared purpose through to the decision.
- ImprovedApprovals
Approvals are now decision briefs
Approvers see what the agent wants to do, why it was held and what happens if they approve — enough to decide without opening another tool.
- SecurityRuntime
Destructive and exfiltration intent governed
Requests to delete, bulk-export or move business data out are recognised as intent, not just keywords, and handled by policy.
- NewMCP
Microsoft Agent Governance Toolkit as a decision input
When enabled, the toolkit's policy engine and MCP scanner can strengthen a verdict to approval, rate-limit or block. It can never weaken an Intertrace decision.
- NewIntegrations
Per-agent integration permissions
A self-hosted token broker holds third-party credentials, and each agent gets only the integrations it's allowed to use.
- SecurityApprovals
Approvals are exact execution contracts
An approval covers exactly the action and arguments that were reviewed. It's consumed into one execution permit and honours the allowed-tools list.
- ImprovedMCP
Execution permits in Monitor mode
MCP calls get permits even while you're still monitoring, so switching an asset to Enforce changes nothing but the verdict.
- NewIntegrations
Devin support for MCP
Devin can connect to both the product MCP and MCP servers behind Intertrace, with sign-in handled through Intertrace.
- NewPlatform
Ask Intertrace, rebuilt
A docked assistant panel that answers questions about your estate and findings in markdown, with real feedback.
- NewReports
Scheduled executive reports
Reports can be scheduled and compared with the previous period, so you can see what changed at a glance.
- ImprovedGuardian
Evidence-based risk calibration
Guardian calibrates risk from evidence and tracks its own false-positive rate, so noisy signals get quieter over time.
- NewRemediation
Tickets that close themselves — with proof
Jira tickets are verified, inbound webhooks keep status in sync, and tickets auto-close only when the fix is shown to be effective.
- NewMCP
MCP posture recommendations
Guardian turns your MCP inventory into specific recommendations you can apply in one click, and tool lists sync on a schedule.
+ The AI watching your AI
See the latest on your own agents.
Book a 30-minute risk assessment, or open the Simulation Lab and try it now.