Robinhood's AI agents can now spend your money

Simon Willison's Blog

OpenAI and Anthropic achieved product-market fit in April 2026 by shifting enterprise customers from discounted subscriptions to full API pricing for coding agents, with evidence showing companies are willing to pay despite budget overruns. This represents a critical inflection point where AI labs can finally generate revenue potentially sufficient to cover their massive infrastructure costs through enterprise adoption of coding agents.

OpenAI

OpenAI's Secure MCP Tunnel enables ChatGPT and other OpenAI products to securely access private MCP servers behind firewalls using outbound-only connections, eliminating the need to expose internal infrastructure to the public internet. This solution addresses a key enterprise adoption barrier by preserving existing network security boundaries while enabling AI integrations.

Robinhood is enabling AI agents to trade stocks and make payments for users through dedicated wallets and virtual credit cards with built-in safeguards. The company joins major tech players in building financial infrastructure for autonomous AI agents.

Anthropic

Anthropic reveals how it secures Claude agents through environmental containment (sandboxes, VMs, egress controls) rather than relying on human oversight or model-layer defenses, sharing critical security incidents including successful credential exfiltration and data leakage that demonstrate why deterministic boundaries are essential as agent capabilities expand.

AgingBench measures how deployed AI agents degrade over time through four distinct mechanisms (compression, interference, revision, and maintenance aging), showing that reliability is a lifespan property requiring mechanism-specific diagnosis rather than snapshot evaluation. Across 14 models and 400+ runs, the research reveals that the same surface failure can stem from different memory pipeline stages, demanding targeted repairs instead of generic fixes.