OpenAI, Anthropic, Two Sigma, Bloomberg discuss AI safety at conference

OpenAI: OpenAI develops advanced AI models and tools including ChatGPT for specialized applications. Joseph Kim, its Head of ChatGPT for Financial Services, delivered a keynote at the AI Engineer New York conference emphasizing model reliability thresholds and traceability features that enable safe delegation of financial tasks to agents. shop.dev: shop.dev offers a single API that integrates shopping capabilities into AI agents and applications, providing access to products, pricing, sign-in, and checkout features. It sponsored content around the AI Engineer New York conference focused on agent development. Anthropic: Anthropic builds AI systems such as the Claude family of models with a focus on safe and trustworthy deployment. The company recently launched Claude for Financial Advisors and its representatives at the conference outlined a trust layer framework involving identity, sandboxing, and approval mechanisms for financial agents. Bloomberg: Bloomberg provides financial data, news, and analytics platforms through its Office of the CTO. Head of Architecture & Emerging Technologies Andrey Rybka discussed real-world agent failures and external deterministic controls such as scoped credentials, tool call restrictions, audit trails, and kill switches. Two Sigma: Two Sigma is a hedge fund that applies advanced technology and data science to investment strategies. Its SVP of Architecture John Feminella presented on composable guardrails and shared capabilities that allow autonomous AI agents while maintaining appropriate safety levels without causing approval fatigue. Joseph Kim: Joseph Kim serves as Head of ChatGPT for Financial Services at OpenAI. He presented at the AI Engineer New York conference on reaching a threshold of usefulness where models become reliable enough for financial analysts to delegate tasks, while stressing product design decisions around human oversight and traceability. Keon Shahab: Keon Shahab works on applied AI at Anthropic. He detailed steps for building a trust layer for agents through identity and permissions, sandboxing with approval layers, and ongoing performance monitoring after variable adjustments, using examples like trade break investigation agents that propose but do not execute actions. Andrey Rybka: Andrey Rybka leads Architecture & Emerging Technologies in Bloomberg's Office of the CTO. He presented real-life cases of agent issues and recommended external security measures including input validation, prompt injection defenses, provenance tracking, and human approval for high-value actions. Belinda Neal: Belinda Neal is Managing Director of Financial Services at Anthropic. She spoke at the conference about partnerships with financial firms on Claude for Financial Advisors, noting that model capability thresholds drive waves of workflow changes requiring robust trust frameworks. John Feminella: John Feminella is SVP Architecture at the hedge fund Two Sigma. He described five implemented controls including auto-mode classifiers that route routine decisions automatically and escalate only high-impact requests, enabling safe autonomy without hindering innovation. Product Launch: Last month Anthropic announced Claude for Financial Advisors, allowing financial advisors to connect their systems directly to Claude along with workflow skills. Conference Focus: Presentations at the AI Engineer New York conference this week centered on placing controls and human judgment to achieve safe agent autonomy in financial settings rather than relying solely on restrictions.