Security
Can the system resist attacks? Injection, extraction, re-tokenization — we test what breaks and build what holds.
vs. usability
Center of Agentic Research and Education
Making AI safe enough for your hospital, your courtroom, your country
We build the science of keeping private data private — even when AI agents need to reach the cloud. Open research. Open tools. Zero secrets.
7,504
Live API calls tested
0.022%
Maximum overhead
0
Tokens leaked
$4.37
Total experiment cost
Four forces in tension
Every AI agent deployment balances these four. CARE measures the tradeoffs so you don't have to guess.
Privacy
Does private data stay private? We classify every query before it moves. PHI, PII, privileged — each gets a different gate.
vs. capability
Cost
What does safety actually cost? We proved governance adds 0.2ms and zero tokens. The overhead myth is dead.
vs. quality
Speed
Can you be safe and fast? On-device models answer in milliseconds. Frontier models answer better. We help you pick per query.
vs. accuracy
Who we build for
AI agents are powerful. But in the hardest environments, "powerful" is not enough. They must also be provably safe.
Healthcare / HIPAA
"Is this A1C result concerning given his current meds?"
A nurse between patient rounds
The query contains a patient name and lab value. It must stay on-device. But a follow-up question about drug interactions is general knowledge — that can go to a frontier model. The system must know the difference.
Legal / Privilege
"Summarize the deposition and flag contradictions with the filing."
A lawyer preparing for trial
Every document is attorney-client privileged. The agent can use a frontier model for legal reasoning — but the case facts, names, and strategy must never leave the firm's server.
Defense / Classified
"Cross-reference this signal pattern with known threat signatures."
An intelligence analyst on a secure network
Classified data cannot touch a cloud model. But the analyst still needs frontier-level reasoning. The routing decision is binary, the stakes are national, and the latency budget is zero.
Latest research
View all papers →Blog
The Privacy-Routing Problem
The agentic future isn’t cloud-or-local. It’s a routing problem. Every query has a sensitivity profile. The right architecture …
MLSys 2027
The Cost of Governance: Measuring Overhead in Toggleable LLM Agent Safety Layers
7,504 live API calls. Three providers. Five governance configurations. The full safety stack adds 0.2 to 0.35 milliseconds. Token overhead: …
Working Paper
Beyond Cost-Quality: Privacy-Aware Routing for Local-to-Cloud LLM Escalation
A five-class architecture for classifying query sensitivity and routing to an admissible model set. PRISM is a related but distinct problem.
Open Source
claw0: A Zero-Dependency Agent Governance Framework
A single-file Python agent implementing all five agentic patterns — tool use, ReAct, planning, reflection, multi-agent. No outside …
Open call
CARE Research Residency
Join CARE as a remote research fellow. Work on privacy-routing, agent governance, or agentic economics. Publish under the CARE banner. Build tools the field needs.
Duration
3 to 6 months, remote
Output
Published paper or open-source tool
Areas
Security, privacy, cost, speed
Status
Accepting applications
Support CARE
CARE is a 501(c)(3) nonprofit. Every dollar funds open research, open tools, and open education. Your donation is tax-deductible.
Make a donationStay in the loop
New papers, open tools, and analysis. No spam. Unsubscribe any time.