OpenSecura
OpenAIDR · AI Detection and Response, by OpenSecura
Autonomous AI Security.
OpenAIDR watches the AI agents on your company’s computers, turns your policies into guardrails, blocks unsafe actions before they run, and hands your team the evidence. Most of it happens without a person in the loop.
Explore more for
to deploy across your company
of AI apps, MCP servers, skills, and packages
of risky configurations identified
accuracy
to detect & contain sophisticated attacks
Built in the open: OpenAIDR is Apache 2.0 open source you can inspect, deploy and extend. Explore the OpenSource page · Research
Let autonomy do the heavy lifting
OpenAIDR brings discovery, posture checks, guardrails, and detection and response together. It handles routine security work and gives your team the evidence and control to act when needed.
Rolls out across every endpoint from one enrollment link, with zero manual configuration.
No professional services, no tuning. Protection starts the day the enrollment link goes out, at any agent count.
Inventories every AI tool, agent, model, MCP server, and API key, and records each agent’s prompts, tool calls, commands, and data access.
Every agent and the actions it takes, including the shadow agents nobody registered.
Identifies risky configurations and excessive permissions across agents and tools.
Will close risky attack paths automatically before exploitation.
Generates adaptive guardrails from a stated intent and enforces them in real time.
Our proprietary models target 99.9% accuracy at one-tenth the cost of competitors.
Detects malicious activity such as prompt injection and data exfiltration, and contains incidents with full forensic evidence.
Risky actions are blocked inline before execution, and an AI investigator contains incidents within minutes.
AI activity map
Swipe sideways to see the whole map →
What OpenAIDR looks for
Agent integrity
Unapproved agents run outside your oversight, or trusted agents are compromised and begin acting against your interests.
See the law firm workflowIdentity & access
Agents use exposed, overprivileged, or unmanaged credentials to access systems and data beyond their intended scope.
See the financial services workflowSupply chain
Malicious or vulnerable tools, MCP servers, skills, and dependencies introduce risk through the integrations your agents rely on.
See the technology workflowAdversarial inputs
Malicious instructions in documents, websites, or tool outputs redirect an agent away from its intended task.
See the healthcare workflowOperational safety
Mistaken or overly broad actions delete data, change critical configurations, or disrupt production services.
See the retail workflowData protection
Agents expose proprietary models, source code, or sensitive business data to unauthorized destinations.
See the manufacturing workflowBuilt for your size of business
Small & Medium Enterprises
OpenSecura runs the hosted service for you: guided onboarding, autonomous operations and a pay-as-you-go subscription. Sensitive data, production services and agent activity stay protected even without a dedicated security team.
Large Enterprises
Enterprise policies become accurate guardrails, alert triage and investigation run automatically, unsafe actions are blocked before damage occurs, and the open-source platform is yours to customize, so AI adoption scales with less manual work.
Ready to try now?
See how OpenAIDR can take on day-to-day AI security work while your team keeps the business moving.
Solutions · Small & Medium Enterprises
Your business relies on AI to get work done. OpenAIDR protects sensitive data, production services, and agent activity—even when you do not have a dedicated security team.
Designed to reduce time spent on AI security by at least 10×—so your team can focus on the business.
WHY OPENAIDR
Get protected with OpenSecura’s hosted service. Guided onboarding, autonomous operations, and flexible subscriptions keep AI security manageable for your team.
Connect your endpoints with self-serve onboarding. OpenSecura runs the service so your team can focus on protecting the business.
See a Demo ↗OpenAIDR accurately detects risks and prevents harmful actions before they damage your business. Give your team the confidence to get more from AI, without adding security worries to their day.
Use a pay-as-you-go subscription as your AI usage grows. Keep onboarding and offboarding straightforward.
Talk to us about hosted service ↗Ready to try now?
See how OpenAIDR can take on day-to-day AI security work while your team keeps the business moving.
Solutions · Large Enterprises
OpenAIDR turns enterprise policies into accurate guardrails, automates alert triage and investigation, and blocks unsafe actions before damage occurs. Scale AI adoption with less manual work and faster protection.
More protection. Less manual work. Built for enterprise scale.
WHY OPENAIDR
Keep pace with your enterprise’s AI adoption. OpenAIDR runs discovery and triage for you, acts at AI speed, and is open source—so your teams shape it to the enterprise instead of waiting on a vendor.
OpenAIDR discovers every approved and unapproved agent, MCP server, skill, and tool, keeps the inventory current, and uses rich context to triage alerts automatically, so your teams review real threats instead of thousands of false alarms.
Automatically detect harmful AI behavior, prevent unsafe actions, and contain attacks at AI speed. OpenAIDR acts before a manual response can catch up, protecting sensitive data and production systems.
OpenAIDR is open source, so your teams can fit it to your own tools, processes, and policies on your schedule, backed by the community and by OpenSecura, instead of waiting months on a vendor roadmap.
Ready when your team is
Bring your toughest attack scenarios. See how OpenAIDR protects your environment and accelerates response.
USE CASES
AI is changing how your business works. OpenAIDR helps protect the data, systems, and relationships that make it yours.
Choose an industry to explore a workflow.
Client confidentiality
AI helps lawyers compare contracts, review discovery, and prepare drafts. Every new assistant or document tool can also become a route out for confidential client information.
A useful-looking skill or MCP server may send entire documents to an outside service. Client-specific restrictions are difficult to enforce consistently across every lawyer, laptop, and agent.
IN PRACTICEExample workflow
A lawyer asks an agent to compare two confidential merger drafts.
The agent selects an unapproved MCP tool that would upload both documents.
The transfer is denied under the client's sharing restrictions; the reason and evidence are recorded.
Protect client information while keeping approved document analysis productive.
OPENAIDR IN ACTION
Sensitive patient information
Administrative assistants summarize records, assemble referral information, and prepare billing documents. These workflows put sensitive patient information within reach of agents and connected tools.
An agent can gather more information than its task requires or send patient records to an unapproved service. A malicious instruction in an incoming document can turn a routine workflow into data exposure.
IN PRACTICEExample workflow
An administrative assistant uses an agent to prepare a referral summary.
A document instruction redirects the agent toward an external service with the patient's full record.
The prohibited transfer is blocked and the related actions are connected for investigation.
Keep useful administrative AI workflows within the organization's data-handling policies.
Customer data and critical systems
Agents help teams investigate transactions, prepare reports, and work across customer and financial records. Their access can extend from reading information to changing business-critical systems.
Credentials and connected tools can give an agent more reach than its task needs. A series of ordinary-looking searches and exports can expose a large volume of confidential information before a person reviews an alert.
IN PRACTICEExample workflow
An analyst asks an agent to summarize a small set of customer transactions.
The agent broadens its search, collects unrelated records, and prepares an external upload.
Connected activity reveals the unsafe path; the prohibited upload is blocked and the findings are reported.
Support faster analysis while protecting customer information and enforcing the boundaries of the task.
Code, credentials, and production
Coding agents read repositories, install dependencies, call MCP tools, and execute commands. Their productivity comes with access to source code, secrets, and production services.
A repository instruction, risky dependency, or over-permissive configuration can redirect an agent. A routine cleanup can become a destructive command, and manual review may arrive after execution.
IN PRACTICEExample workflow
An operations agent is asked to remove temporary development data.
Its proposed command targets production records instead of the test environment.
Context-aware guardrails block the destructive action before execution and return the policy reason.
Preserve development speed while reducing the risk of exposed secrets and preventable outages.
Customer trust and operations
AI assistants answer customer questions, work with order histories, and help teams maintain catalogs and inventory. One workflow can touch several systems and thousands of customer records.
An agent may expose customer information, follow instructions hidden in a support request, or make unauthorized changes across many records. A flood of alerts makes the meaningful incidents harder to find.
IN PRACTICEExample workflow
An agent investigates a customer's missing order.
Instructions in the request push it to retrieve unrelated customer records and send them outside approved systems.
The unsafe activity is identified, the prohibited transfer is stopped, and the connected actions are reported.
Keep customer-facing AI useful while protecting customer data and operational systems.
Intellectual property and continuity
Engineering and procurement teams use agents to review specifications, compare supplier proposals, and prepare operational updates. These tasks bring proprietary designs and internal systems into the same workflow.
An external supplier document can carry a hidden instruction. An agent with broad access may then gather confidential designs or attempt an unauthorized change to an operational business system.
IN PRACTICEExample workflow
An engineer asks an agent to compare a supplier proposal with internal specifications.
A hidden instruction prompts the agent to gather additional design files and upload them externally.
The sequence is identified as a connected threat; the prohibited upload is blocked and evidence is retained.
Use AI to accelerate engineering and procurement while protecting proprietary information.
YOUR WORKFLOW. YOUR POLICIES.
See how OpenAIDR can protect the AI workflows your business depends on.
Research · MLSys 2026
The paper introduces ADR, an enterprise security system developed at Uber to detect threats in AI agent activity. It combines detailed telemetry of prompts, reasoning, and tool use with fast triage, context-aware investigation, and offline red teaming.
Evaluated in Uber's production environment and on research benchmarks, ADR shows how teams can investigate unsafe agent behavior while keeping detection costs manageable at scale.
Chenning Li et al. · MLSys 2026 · Industry Track
OpenSource / OpenAIDR
For anyone who wants a solid AI security foundation in their own environment. Start with visibility, add your guardrails, and build on code you can inspect and customize.
The story behind the project
OpenAIDR was built by major contributors to Uber’s open-source ADR project. After leaving Uber, they co-founded OpenSecura to bring their experience securing enterprise AI to a platform anyone can deploy and extend.
That experience shapes a practical goal: make AI security useful in your environment, without requiring an existing security operations stack. The source is Apache 2.0: self-host the backend, connect your endpoints, and adapt it to your needs.
Explore the sourceWhat is new in OpenAIDR
Discover
See which AI agents are running, what they access, and which risky configurations need attention—without building the inventory by hand.
Generate
State what your business needs to protect. OpenAIDR turns that intent into adaptive guardrails that prevent unsafe actions while allowing legitimate work.
Investigate
Connect seemingly harmless actions into attack paths. OpenAIDR brings the evidence together, so your team spends less time chasing noise and reconstructing incidents.
Defend
Block data leaks and other unsafe actions before they run. Contain malicious attempts automatically, keeping your business protected while your team stays focused.
The open-source edition includes deployment, discovery, posture checks, and manual guardrails. Adaptive guardrails, automated remediation, and the AI investigator are OpenSecura’s autonomous features, powered by its proprietary models.
Deployment
Install the desktop app for macOS, Windows, or Linux and choose Continue for free. No account, API key, or server is needed, and every install includes a free allowance of cloud guardrail checks.
Evaluate the full workspace on one computer. In the app, choose Connect organization → Set up local runtime; Docker creates the database, server, and dashboard and enrolls the device. For a source build or another platform, use Compose:
git clone https://github.com/opensecurainc/openaidr.git
cd openaidr
python3 scripts/local_stack.py configure
docker compose up --build -dOpen http://localhost:8080 and sign in as admin@localhost with the generated OPENAIDR_ADMIN_PASSWORD in .env.
Compose is for development and evaluation, not production. Before enrolling other computers, set the public URL and bind address behind your own HTTPS reverse proxy.
Connect devices to a workspace: your local runtime, or a managed workspace from OpenSecura.
Engineering, research, and lessons in securing AI.
Product · November 2026
Apply one policy across supported agents, see the tools they use, and investigate risky actions. Try OpenAIDR for free without creating an account.
Read the article OpenSecura team · 2 min read
OpenSecura · Product
Visibility and guardrails for AI coding agents.
An AI coding agent is about to read your SSH keys or run a destructive command. With a matching rule and Enforce mode enabled, OpenAIDR can intervene before the action runs at a supported agent hook.
OpenAIDR by OpenSecura gives security teams visibility and control over what AI agents actually do:
Start in Shadow mode to evaluate guardrail decisions while work continues. Enable Enforce when you are ready to apply them. Integrations include Claude Code, Codex, Cursor, and GitHub Copilot, with coverage varying by agent.
See supported integrations and guardrail behavior.
The desktop app, sensor, dashboard, and local server are Apache 2.0 licensed, so you can inspect the code and run the core yourself. Model-based security checks use OpenSecura’s hosted Intelligence service. You can also track token usage and estimated costs.
Read about the open-source boundary.
OpenAIDR is under active development. Share feedback or contribute on GitHub.
Our mission
As AI becomes part of everyday work, security must evolve alongside it.
Protect AI systems and agent workflows so organizations can adopt them with confidence.
Help security teams use AI to understand evidence, investigate risk, and respond effectively.
Build defenses for threats that exploit AI capabilities, autonomy, and access to real systems.
Our story
We are starting with a concrete challenge: understanding and protecting the work AI agents do.
Founded with a mission to secure AI, secure with AI, and secure against AI.
Our first focus
With OpenAIDR, we are bringing our mission to the agent workflow: observe activity, detect unsafe behavior, give teams evidence to respond, and prevent unsafe actions.
Explore the solution