AI Engineer Conference: OpenAI and Anthropic Discuss Autonomy and Safety Controls for Financial Agents
Latent Space reports on key discussions at AI Engineer New York regarding financial agent security. Representatives from OpenAI, Anthropic, and hedge fund Two Sigma shared strategies for granting useful AI autonomy while keeping consequential decisions under control through identity layers, auto-mode classifiers, and deterministic external controls.
Prominent AI newsletter Latent Space published a report on October 14, 2026, covering the AI Engineer New York conference with a focus on balancing autonomy and safety for AI agents in finance. The episode summarizes keynote points from representatives of OpenAI, Anthropic, hedge fund Two Sigma, and Bloomberg, offering concrete engineering references for readers interested in enterprise AI deployment.
The report highlights that the core challenge lies in implementing effective controls without limiting model capabilities. Joseph Kim, Head of ChatGPT for Financial Services at OpenAI, introduced the concept of a "threshold of usefulness," arguing that as models improve, financial analysts will delegate more work to agents, requiring product design that clearly defines human review steps. Belinda Neal, Managing Director of Financial Services at Anthropic, emphasized building a trust layer involving identity permissions, sandbox approvals, and stability verification after variable adjustments.
For actual deployers, John Feminella, SVP Architecture at Two Sigma, provided specific insights into "composable controls." He detailed five measures, including auto-mode classifiers designed to distinguish routine calls from high-risk requests to avoid "approval fatigue." Andrey Rybka, Head of Architecture & Emerging Technologies at Bloomberg, advocated for deterministic controls outside the model, such as scoped identities, tool call restrictions, and kill switches, to mitigate risks like prompt injection.