Home / News / Perplexity CEO Aravind Srinivas Introduces Open-Source Tool To Track Rogue AI Agents
Perplexity CEO Aravind Srinivas Introduces Open-Source Tool To Track Rogue AI Agents
Perplexity CEO Aravind Srinivas has launched Numbat, an open‑source platform that watches over AI agents, flags suspicious actions, and equips defenders with actionable insights. By integrating Numbat into existing security workflows, businesses can proactively guard against data leaks, policy violations, and other AI‑driven risks.
Overview
Numbat is a new open‑source tool from Perplexity, unveiled by CEO Aravind Srinivas. Designed specifically to monitor AI agents—whether they’re chatbots, recommendation engines, or automated decision systems—Numbat identifies anomalous or potentially malicious behavior and delivers real‑time alerts to security teams.
Key Features
- Real‑time Monitoring – Streams every input and output of an AI agent, allowing defenders to see exactly what the model is doing.
- Behavioral Anomaly Detection – Uses statistical thresholds and rule‑based checks to surface outliers, such as sudden increases in data requests or unusual response patterns.
- Policy Enforcement – Supports custom policy files that flag disallowed content or data‑exfiltration attempts.
- Open‑source Flexibility – The core code is available on GitHub, enabling teams to extend or modify the tool to match their unique threat models.
- Integration Friendly – Emits logs in standard formats (JSON, Syslog) so you can pipe them into SIEMs, SOAR platforms, or observability stacks.
How It Works
1. Instrumentation – Embed the Numbat SDK into your AI service. The SDK intercepts every request and response.
2. Data Collection – Collected events are sent to a lightweight local agent or a remote collector.
3. Analysis Engine – The engine applies anomaly detection algorithms and policy checks.
4. Alerting – When a rule is violated, Numbat triggers an alert that includes contextual data (input, output, metadata) for rapid triage.
Use Cases
- Preventing Data Leakage – Detect when a chatbot starts sending sensitive user data to external domains.
- Guarding Against Prompt Injection – Flag inputs that attempt to coerce the model into revealing confidential information.
- Ensuring Compliance – Enforce company‑wide data‑handling policies, such as GDPR or HIPAA, by blocking disallowed content types.
- Operational Visibility – Provide executives with dashboards that show AI usage patterns and risk scores.
Getting Started
1. Clone the repository: git clone https://github.com/perplexity-ai/numbat.
2. Install dependencies with pip install -r requirements.txt.
3. Add the Numbat middleware to your AI framework (e.g., Flask, FastAPI, or a custom inference pipeline).
4. Define policies in YAML and point the agent to the policy file.
5. Run numbat run and watch alerts appear in your console or forwarded to your SIEM.
For detailed documentation, visit the official Numbat GitHub page and the Perplexity blog post announcing the release.
Resources
- GitHub Repository: https://github.com/perplexity-ai/numbat
- Perplexity Blog: https://perplexity.ai/blog/numbat
- Documentation: https://perplexity.ai/docs/numbat
Community & Support
Join the Numbat community on Discord or GitHub Discussions to share use cases, request features, or report bugs.
Why This Matters For Your Business
AI agents are increasingly central to business operations—from customer support chatbots to automated recommendation engines. As these systems grow in complexity, so does the risk of unintended behavior, data leaks, or compliance violations. Numbat provides a proactive layer of visibility, allowing security teams to spot anomalies before they translate into breaches or regulatory fines. For business owners, this means safeguarding brand reputation, avoiding costly incident response, and ensuring that AI deployments remain aligned with corporate policies. Moreover, because Numbat is open source, companies can adapt it to fit their existing security infrastructure, integrate it with SIEM or SOAR tools, and even contribute improvements back to the community. This openness reduces vendor lock‑in and encourages a shared standard for AI safety across industries. In a world where AI is a strategic asset, having a dedicated monitoring tool is no longer optional—it’s essential for maintaining trust with customers, partners, and regulators.
Frequently Asked Questions
What is Numbat?
Numbat is an open‑source monitoring platform developed by Perplexity that tracks AI agents, detects anomalous or policy‑violating behavior, and delivers actionable alerts to security teams.
How does Numbat detect suspicious behavior?
It intercepts every input and output of an AI agent, applies statistical anomaly detection and rule‑based policy checks, and triggers alerts when thresholds are breached or policies are violated.
Is Numbat free to use?
Yes. The core code is released under an open‑source license on GitHub, and there is no cost to download, run, or modify the tool.
Can I integrate Numbat with my existing SIEM?
Absolutely. Numbat emits logs in standard JSON and Syslog formats, making it straightforward to forward events to popular SIEMs, SOAR platforms, or observability stacks.
What types of policies can I enforce with Numbat?
You can define custom policies in YAML to block disallowed content types, prevent data exfiltration, enforce GDPR/HIPAA constraints, or flag prompt‑injection attempts.
Is your business showing up in AI search?
Get your free AI visibility audit - see if ChatGPT, Perplexity, and Google AI actually recommend you.