University of Pennsylvania Pilots PennChat: A Secure AI Portal Powered by Claude and ChatGPT

What PennChat Actually Is

PennChat is built on LibreChat, an open-source platform, and hosted on Amazon Web Services within Penn’s secure network. Anthropic’s Claude models are accessed through Amazon Bedrock. OpenAI’s ChatGPT models come through Microsoft Foundry.

The key detail: neither provider can use Penn data to train their models. That’s a non-negotiable for any institution handling sensitive research, student records, or proprietary data.

Users get access to 14 models sorted into three tiers — premium, balanced, and economical — plus a legacy category for older LLMs. Each user receives a daily credit balance to manage usage across those models.

The Security Architecture Is the Story

Penn didn’t just flip on an API key and call it a day. The setup reflects deliberate choices:

• Network-gated access — users must be on PennNet, AirPennNet, or the University’s GlobalProtect VPN
• No live internet queries — LLMs cannot reach the open web, reducing data leakage risk
• Data classification alignment — the platform is cleared for Low, Moderate, and most High Risk data per Penn’s own classification framework
• Restricted admin access — only a small group of ISC administrators can view chat history, and only for troubleshooting

That last point matters for trust. Hanulec stated publicly that they have “no intention of looking at that data” outside of user-requested support. Transparency about who can see what is often missing from enterprise AI rollouts.

What Users Can Do With It

Beyond basic prompting, PennChat includes features that push it closer to a real productivity platform:

• Custom agent builder — select a model, write instructions, toggle output parameters
• Project folders — organize multiple chats by topic or workflow
• File uploads and document generation — supports Word, Excel, CSV, and PDF output

The ability to build custom agents is particularly notable for research teams and administrative departments that need repeatable, specialized workflows rather than one-off queries.

The Data Caveats Worth Knowing

PennChat’s LLMs have training data cutoffs ranging from 2024 to 2025. ISC recommends asking the specific model you’re using for its exact cutoff — a practical tip that most enterprise AI deployments don’t bother surfacing to users.

Despite the platform’s high-risk data clearance, Penn explicitly advises against entering Social Security numbers, credit card numbers, or protected health information. Cleared for high-risk doesn’t mean cleared for everything.

Why This Matters Beyond Penn’s Campus

Penn isn’t alone here. Dartmouth and the University of Chicago have launched similar institutional LLM platforms recently. The pattern is consistent: universities are moving from ad-hoc ChatGPT use to structured, governed, institution-wide deployments.

For AI tool observers, this is a meaningful shift. It signals:

• Demand for enterprise-grade AI wrappers that sit between raw APIs and end users
• Growing preference for multi-model access rather than single-vendor lock-in
• Data privacy as a baseline requirement, not a differentiator

Open-source platforms like LibreChat are becoming the connective tissue for these deployments — flexible enough to customize, credible enough to trust.

What’s Coming Next

Penn has flagged two areas of future development: agent sharing between users and connections to external data sources like PubMed. The PubMed integration would be significant for research use cases — giving faculty and students grounded, domain-specific retrieval on top of general LLM capabilities.

Full launch is expected in August, at which point Penn plans to publish clearer guidance on cost, capability, and data sensitivity levels for each available model.

The Practical Takeaway

If you’re evaluating AI tools for an institution, enterprise, or organization with real data governance requirements, PennChat’s architecture is worth studying. The combination of open-source infrastructure, multi-cloud model access, network-gated security, and tiered data classification is a replicable blueprint — not a one-off academic experiment.

The question isn’t whether your organization needs something like this. It’s whether you’re building it deliberately or letting shadow AI use make the decision for you.