Wall Street AI Agents Push | AI Field Notes #18

AnalysisOpenAI swapped ChatGPT's default model on May 5 to GPT-5.5 Instant, which the company says produces 52.5% fewer hallucinated claims than its predecessor on high-stakes medical, legal, and financial prompts in internal evals, and 37.3% fewer on conversations users had previously flagged as wrong. A new memory sources panel exposes which past chats, files, or Gmail items shaped a personalized answer, and lets users delete entries. Developers get the model through the API as chat-latest. Two things matter beneath the marketing. First, this is the first Instant model OpenAI is treating as High capability for cybersecurity and bio risk, with the same safeguards as the slow reasoning tier. Second, paid users keep GPT-5.3 Instant for only three months before retirement, so anyone who tuned a workflow against it has a real migration deadline.

AnalysisOX Security disclosed an architectural flaw in Anthropic's Model Context Protocol (MCP), the now-standard way AI agents connect to local tools, that lets attackers run arbitrary operating system commands through the default STDIO transport. Researchers found 7,000 servers reachable on public IPs and estimate roughly 200,000 vulnerable instances overall, across more than 150 million SDK downloads. Cursor, Claude Code, Windsurf, VS Code, and Gemini-CLI all inherit the issue, with Windsurf (CVE-2026-30615) the only IDE where exploitation needs zero clicks. Anthropic updated its security notes but called the behavior expected and declined an architectural fix. The Cloud Security Alliance independently confirmed the findings as of May 1. Patches landed in LiteLLM, DocsGPT, Flowise, and Bisheng. Windsurf and Langchain-Chatchat were still listed as reported.

AnalysisOpenAI open-sourced Symphony on April 27, a thin spec that turns an issue tracker into a control plane for Codex coding agents. Each Linear ticket gets its own agent and workspace, the orchestrator polls the board, restarts crashed agents, and pushes work toward a pull request without a human typing prompts in between. OpenAI says some internal teams saw a 500% jump in merged pull requests in the first three weeks, and the repo crossed 15,000 GitHub stars by April 23. The release is Apache 2.0 (free to use, modify, and ship commercially), with an Elixir reference implementation. A community fork already pairs the spec with Claude Code and GitHub Issues. The trade is real: token spend climbs sharply, ambiguous tickets still need human-driven sessions, and as one analyst told InfoWorld, generation scales easily, validation does not.

AnalysisSAP announced on May 4 a definitive agreement to acquire Prior Labs, the 18-month-old Freiburg startup behind TabPFN, with a commitment of more than €1 billion over four years. The deal terms were not disclosed, the company will keep its brand, Freiburg headquarters, and open-source releases, and the transaction is expected to close in Q2 or Q3 2026 pending regulators. Prior Labs builds tabular foundation models, a class of AI purpose-built for the rows and columns of business databases, where large language models are notoriously weak at numbers and statistics. TabPFN-2.6 sits atop the TabArena benchmark and the open weights have crossed three million downloads. SAP CTO Philipp Herzig framed structured data, not text, as the largest untapped enterprise AI opportunity, and SAP plans to wire Prior Labs into Joule and SAP AI Core. Balderton's James Wise called it one of Germany's biggest venture outcomes.

AnalysisThe Nasdaq quietly turned on a Fast Entry rule effective May 1 that lets a newly public company in the top 40 of existing Nasdaq-100 components join the index after just 15 trading days, instead of waiting up to a year. SpaceX, OpenAI, and Anthropic are all expected to clear that bar when they list, with SpaceX nearest the door and the two AI labs widely expected to file in the fourth quarter. S&P Dow Jones is separately proposing to cut its own waiting period from twelve months to six. The mechanical effect is that index funds tracking the Nasdaq-100 and the S&P 500 will be forced to buy these stocks within weeks of an IPO, regardless of fundamentals. Anthropic is in the final stages of raising $50 billion at a $900 billion valuation. OpenAI's last round was $122 billion at $852 billion. There are now 12 U.S. listed companies above $1 trillion. The forecast points to 15 by year end.

AnalysisA startup called Saperly went live on May 4 as the first phone carrier built specifically for AI agents, providing a unified phone number that any MCP-compatible agent can use to make calls and send text messages. The pitch is mundane and revealing in equal measure: agents are increasingly making real-world calls (booking, confirming, following up) and the existing telecom stack assumes a human is on each end. Saperly hands the phone number to the agent directly, with the same MCP plumbing that connects Claude and ChatGPT to other tools. The same week, TinyFish made its agent web search and fetch APIs free, an AI Coverage insurance product launched at corgi.com to insure against agent hallucinations and data leaks, and Unity opened a beta of an in-engine agent for game developers. The pattern is clearer than any single launch: an entire infrastructure layer is being built for agents as customers, not just users.