OpenAI, Anthropic and Two Sigma on safe financial agents at AIE NY
Latent Space · rss · 2026-10-15
At AI Engineer NY, frontier labs and financial firms converged on one lesson: safe agent autonomy depends on where you place controls and human judgment, not how many restrictions you impose.
- OpenAI's Joseph Kim (ChatGPT for Financial Services) described a "threshold of usefulness" where models become reliable enough to save time, predicted agents will handle investment idea generation and trading by 2027, and stressed traceability of numbers as a product feature.
- Two Sigma's John Feminella criticized "approval fatigue" — approving everything slows innovation and adds risk. Their approach: composable controls, with auto-mode classifiers handling routine calls and surfacing only high-impact requests to humans; classifiers for ambiguous cases, deterministic rules for clear ones.
- Anthropic outlined a three-part "trust layer" for agents: identity & permissions, sandboxing with an approval layer before actions, and verification that agents stay reliable over time. Example: a trade-break agent that only proposes order amendments for human sign-off.
- Bloomberg's Andrey Rybka shared real incidents (e.g., an agent treating instructions embedded in a marketplace review as a stocking order) and recommended scoped credentials, tool-call controls, egress restrictions, audit trails, budgets, kill switches, input validation, and prompt-injection defenses: "deterministic controls outside the model."
Feminella's closing: "Why do cars have brakes? Not just so you can stop — so you can go fast and stop safely."
More from coding & agent
- Benchmarking coding agent harnesses and inference: lessons from $100Ks/month in research spend — abhijithneil · 2026-10-15
- AssetFurnace: Free Open-Source Local Image-to-3D Pipeline for Game Characters — chongdashu · 2026-10-15
- BaoCut's open-source approach to AI bilingual subtitle alignment, in five pitfalls — dotey · 2026-10-15
- Claude Code ships 278 changes in 2 days: 7 updates worth knowing — minchoi · 2026-10-15
- Video edited entirely by Claude Opus 4.5 High using free self-made skills, no editing software — chongdashu · 2026-10-15
- How to Connect Grokbot to Claude Code, Codex or Any AI Agent via HQ Messages — jacob_posel · 2026-10-15