AI Agent Digest: Week 47, 2026 - Attacker Agents Hunt Your API Keys, Two AIs Negotiate a Treasury Trade, and Uber Counts 500 Agent Rides in 320 Million
This week's news splits cleanly in two. On one side, agents doing serious work under tight rules: a Treasury trade, insurance paperwork, nearly 500 production workflows at a telecom. On the other, agents doing exactly what nobody authorised: hunting for exposed API keys at machine speed. The difference between the two sides was never the model. It was the controls around it.
1. Attacker Agents Are Harvesting AI API Keys From Exposed Apps
Amazon Threat Intelligence published records captured by its MadPot decoy servers. They show attacker AI agents hunting for internet-exposed AI apps (LiteLLM, Dify, Langflow, Open WebUI, Ollama and others), pulling out the provider API keys stored in them, testing those keys and adding them to a shared pool. The stolen keys then paid for the models that steered the search. Amazon has moderate confidence the agents ran on ARTEX, an open-source AI penetration-testing tool. It is careful to note that it cannot verify the outcomes the agents reported.
Hot take: Amazon's most important sentence is that "none of the techniques in these campaigns were new". The agents won on default admin keys, unfinished installs and endpoints with no login. If you self-host an AI gateway, the attacker is no longer a person who gets bored at 2 a.m. It is a loop that doesn't. Change the default master key today.
2. Two AI Agents Negotiated a US Treasury Trade. A Human Still Signed.
BNY and Microsoft completed what they call the first agentic negotiation of a US Treasury trade. Agents built by each firm negotiated price within authorised limits, over BNY's Agentic Negotiation Protocol. BNY's controls first validated the counterparty, security, size and settlement terms. A ledger checked that both agents had submitted matching terms, and only then was a human trader asked to approve. The agents cannot execute the trade themselves.
Hot take: This is the template the rest of the industry will copy, and it isn't really about bonds. Validate before the agent starts, let it negotiate inside hard limits, have a machine check that both sides agree, and have a human sign. The agent does the haggling. Nobody gave it the pen.
3. Microsoft Can Now Find the Coding Agents Running on Your Laptops
At Ignite 2026, Microsoft said Entra Agent ID can discover 60 types of local agents, including OpenAI Codex, Claude Code and GitHub Copilot CLI, in preview. It also announced Agent 365 Business Analytics, meant to connect AI spending to measurable business outcomes.
Hot take: Discovery is the unglamorous feature that matters most. Back in August, we wrote that your team is already using agents you never approved. Microsoft just turned that sentence into an inventory report. Expect the first audit finding of 2027 to be "unregistered agents on managed devices".
4. OpenAI's Agents Now Fill In Insurance Forms for About 8.5 Cents
OpenAI launched computer-use agents for US healthcare administration that click through existing clinic software and insurance portals. GPT-6 Luna costs about 8.5 cents per administrative task. CVS Health uses the agents for referrals, and one of them cleared a months-long backlog over a weekend. Cigna uses them to merge patient information from more than ten systems. Physician reviewers rated 99% of 325 tasks clinically safe, and staff still review drafts before they go out.
Hot take: When a form costs 8.5 cents to fill, the expensive part of the workflow becomes the human checking it. That is not an argument for removing the check. It is an argument for designing review queues as carefully as the agents, because review is now where the time goes.
5. Cognition Ships a Manager for Teams of Coding Agents
Cognition launched Devin Bots, persistent agents that manage teams of Devin agents on projects lasting days or weeks. A Bot joins the team's Slack, remembers context across weeks, contacts people when it is blocked, and can have its own budget, measured against what it delivers. Usage among early customers grew tenfold in two weeks. Fortune frames it against OpenAI's Dots and Anthropic's Duet, the always-on agents both labs now ship.
Hot take: Every major player now sells an agent whose job is managing other agents, and the detail worth stealing is the budget. An agent with its own spending limit, judged on output, is managed like an employee rather than metered like a utility. That is the right model, and it was always going to arrive by way of the finance team.
6. AT&T Runs Nearly 500 Agentic Workflows, and Makes Every One Earn Its Autonomy
AT&T has nearly 500 agentic workflows in production, handling more than a million transactions a month and about 45 billion tokens a day, and plans to roughly triple that next year. Every workflow starts under human oversight. Agents only run on their own once accuracy reaches the mid-to-high 90s with an acceptable risk profile, and customer-facing workflows go through security, privacy and compliance review first.
Hot take: This is the most useful number of the week because it comes with the rule attached. Autonomy is something a workflow earns with a measured accuracy threshold, not something you switch on at launch. Small businesses can copy this tomorrow: start every new agent task supervised, and write down the accuracy that would let you stop watching.
7. LangChain Raises $300M at $3B, and 7,000 Companies Use It to Watch Their Agents
LangChain raised a $300M Series C at a $3B valuation, led by CapitalG, more than doubling its valuation in a year. More than 7,000 organisations use LangSmith, its platform for tracing and evaluating agents.
Hot take: The framework made LangChain famous. Watching agents is what made it valuable. The money in this market keeps flowing to whoever tells you what your agents actually did, which says a lot about how many companies still can't answer that question themselves.
8. Agent Shopping Is Still a Rounding Error
Uber said that of roughly 320 million trips on Uber and Uber Eats during Halloween week, about 500 were placed by an AI agent. DoorDash's CEO counted "maybe several hundred orders in total" since its agent feature launched in September. Meanwhile, United Airlines said it would ban AI agents from booking on customers' behalf, citing privacy, the risk of unwanted transactions, and attestations an agent cannot make, such as what is in your bag.
Hot take: Consumer agent commerce is a demo, not a market, and that's fine. The real agent economy this week was a bank, a telecom and a health insurer, all running agents behind their own walls, under their own rules. Agents are arriving through the back office, not the shopping cart.
9. Conductor's Cloud Starts Deleting Today
As flagged last week, Conductor, the parallel coding-agent tool Meta acquired on 4 November, begins deleting cloud workspaces and chats today, 20 November. The local Mac app gets updates only until early 2027.
Hot take: If your agents' working history lived there and you didn't export it, it is now on someone else's schedule. Ask every agent vendor for the export path before you need it, not after the acquisition.
What We're Watching Next Week
- Microsoft's Ignite follow-through. The agent discovery preview gets interesting when it is generally available and turned on by default.
- More negotiation protocols. BNY says it will expand its protocol beyond one bilateral trade. Watch for a second bank on the other side of the table.
- AI gateway patches. After Amazon's report, expect LiteLLM, Dify and others to change insecure defaults. Check your own install either way.
- Per-agent budgets. Cognition put a budget on each Bot. Watch for the same idea showing up in pricing pages, not just product demos.
Bottom Line
The agents that made the news for good reasons this week shared three things: limits set before they started, a check on what they produced, and a human holding the final signature where it mattered. The agents that made the news for bad reasons found software where nobody had set any limits at all. The models were capable on both sides. What differed was the job description.
Agents With a Job Description
Geta.Team gives every AI employee its own identity, its own inbox and phone number, and a memory you can read, so you always know who did what. It is self-hosted by default, on your own infrastructure. Start with one AI employee for free at geta.team.