Make the threat concrete.
Specify the sensitive information, the adversary’s access, and the decisions an agent is permitted to make.
AI safety research · Finance first
Intelligence, with
boundaries.
Building the foundations for AI that can work with sensitive financial information and act within verifiable limits.
Explore our researchPrivacy by designAdversarial resilienceControlled autonomy
01 / The mission
Finance is our starting point.
Safer AI is our ambition.
AI can help uncover investment ideas, support financial operations, and coordinate complex decisions. But useful capability brings difficult questions: who sees the data, what can the system execute, and how does it behave under attack? We are building a research lab around those questions.
02 / Research agenda
Five problems.
One safety mandate.
Our current research directions. Open each area to explore the questions guiding our work.
01Private financial intelligence
How can models and agents support hedge fund research, investment hypothesis generation, and factor discovery while protecting proprietary datasets, strategies, research pipelines, and generated alpha?
02Confidential financial operations
We study privacy-preserving AI for accounting, costing, book building, valuation, due diligence, and corporate finance. The research question extends beyond access: can we limit data transmission, retention, and unintended incorporation into training, fine-tuning, or continual learning?
03Adversarial robustness
Financial agents consume news, filings, research, market data, and tool outputs. We investigate defenses against prompt injection, poisoned inputs, compromised APIs, and data exfiltration that could expose strategies or manipulate decisions and trades.
04Trust & secure delegation
What should an agent be allowed to access, remember, communicate, and execute? We explore fine-grained permissions, secure execution, restricted information flows, auditable decision trails, and verifiable action limits for agents connected to accounts and financial infrastructure.
05Safety between agents
As autonomous financial agents interact, local failures may become collective risks. We investigate unintended coordination, algorithmic collusion, manipulation, strategy leakage, and the propagation of errors, with systemic stability as a central concern.
03 / Our approach
From precise questions
to testable safeguards.
Our intended path combines foundational research with experimental implementation. Safety claims should be tied to explicit assumptions and evidence.
Build and investigate.
Study mathematical mechanisms and implement prototypes that put proposed safeguards to the test.
Measure the limits.
Examine failure cases, attack resistance, and the tradeoffs between useful capability and enforceable control.
04 / Contact
Work on the
hard questions.
For research conversations and potential collaboration in AI safety for finance, get in touch.