Crossing the Threshold: How OpenAI and Anthropic Approach Autonomy and Safety in Financial Agents
OpenAI, Anthropic, Two Sigma, and Bloomberg outlined controls for safer autonomous financial agents.
At AI Engineer New York, speakers from OpenAI, Anthropic, Two Sigma, and Bloomberg described how financial firms can grant agents autonomy without removing human control. OpenAI's Joseph Kim said ChatGPT for Financial Services needs traceable numbers and predicted that, by 2027, agents could support investment-idea generation and trading. Anthropic discussed Claude for Financial Advisors and a trust layer covering identity and permissions, sandboxed approval before action, and ongoing evaluation, using a trade-break agent that only proposes order amendments. Bloomberg's Andrey Rybka cited agents following instructions embedded in content and recommended scoped credentials, tool-call controls, traffic restrictions, audit trails, budgets, kill switches, and prompt-injection defenses outside the model.
- OpenAI's Joseph Kim predicted finance agents could generate ideas and trade by 2027.
- Two Sigma uses classifiers so people review only high-impact agent actions.
- Anthropic separates agent identity, human approval, and ongoing reliability checks.
- Bloomberg recommended external controls, audit logs, kill switches, and prompt-injection defenses.
- Tags
- agent-securityopenaianthropicfinancial-servicesprompt-injectiontwo-sigmabloomberghuman-in-the-loop
From the original article
5 min read at latent.space
At AI Engineer New York, frontier labs and financial firms explored how to give agents useful autonomy while keeping consequential decisions under control.
By Richard MacManus · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.