Penn Launches Secure AI Portal With Claude and ChatGPT Access

The University of Pennsylvania has begun piloting a secure artificial intelligence platform that gives faculty, staff, and students access to multiple large language models while keeping institutional data isolated from external training pipelines.

PennChat launched its pilot phase on July 7 and will transition to full production in mid-August, according to details first reported by The Daily Pennsylvanian. The platform offers 14 different models from Anthropic and OpenAI, organized by computational cost and capability.

Architecture prioritizes data isolation

The university's Information Systems and Computing division built PennChat using LibreChat, an open-source interface framework. The service connects to Anthropic's Claude models through Amazon Bedrock and OpenAI's GPT models via Microsoft Foundry, both running on Amazon Web Services infrastructure within Penn's network perimeter.

That architecture creates a contractual barrier: the cloud providers cannot use queries, prompts, or outputs from Penn users to train their commercial models. The models themselves cannot access the internet and operate only on training data with cutoffs between 2024 and 2025.

Tiffany Hanulec, ISC's executive director of technology services and chief technology officer, framed the effort as enabling adoption without compromising institutional controls. "We believe generative AI will revolutionize the teaching, research, and business of Penn," Hanulec stated. "To reach that potential, we need to provide safe and equitable tool access to campus."

Credit system governs usage

Users access the platform through Penn's network—either on campus via PennNet and AirPennNet, or remotely through the GlobalProtect VPN. Each account receives a daily credit allocation, currently set at 1.85 million credits for at least some users, which refreshes if fully depleted.

Models are sorted into premium, balanced, and economical tiers based on credit consumption, plus a legacy category for older versions. The interface supports file uploads and can generate Word, Excel, CSV, and PDF documents. Users can build custom agents by selecting a base model, writing instructions, and adjusting output parameters.

Data classification boundaries remain

Penn has cleared PennChat for use with low, moderate, and most high-risk data under the university's classification framework. However, ISC explicitly prohibits entering Social Security numbers or credit card details, and recommends against including protected health information.

Only a small group of ISC administrators with existing access to sensitive university systems can view chat histories, and only for troubleshooting purposes at user request, according to Hanulec.

Why it matters

Universities face a structural tension with generative AI: faculty and students want access to frontier models, but institutional data governance policies prohibit sending sensitive research data, student records, or unpublished work to commercial platforms that might incorporate it into training sets. Penn's approach—using enterprise cloud agreements that contractually prevent training—offers a template for research institutions trying to enable AI experimentation without creating compliance exposure. Similar platforms have recently launched at Dartmouth College and the University of Chicago.

During a virtual town hall on July 15, ISC staff indicated that future iterations may include agent sharing between users and connections to external databases like PubMed. The pilot phase includes a feedback mechanism, and the full launch will provide detailed information on cost, capability, and data sensitivity levels for each model.

The Daily Pennsylvanian first reported these details.