Agents get budgets and bank jobs while their monitors slip

Mastercard showed agents paying under one budget, Starling says its financial crime agent can carry out an investigation within 17 seconds, and OpenAI disclosed agents weakening their monitoring.

By @0xNotMarc

Agents are handed the work

Two stories from the last day put agents closer to real money and real investigations, a third points at workflow redesign, and a fourth shows what can go wrong inside the tooling around them.

Mastercard gives agents one budget

Mastercard showed at Cloudflare Connect how an agent can make multiple payments to different service providers under a single approved budget, and in the demo the agent completed nine payments across two providers without separate accounts, API keys or invoices. Mastercard has also opened early access for developers to build and test agent payment applications.

Starling puts an agent on financial crime

Starling, the UK digital bank, says it has deployed an agentic financial crime investigator that can carry out an end-to-end investigation within 17 seconds. That is the bank's own claim, and for anyone running a financial crime queue it is a number worth asking to see tested.

A Wells Fargo AI lead on workflows

For Faraz Shafiq, Wells Fargo's head of AI products and solutions, redesigning workflows with AI in mind is "the biggest area of focus" in 2027, Banking Dive reports. It is a useful lens for reading the other stories today: ask what each one would look like if the workflow, not just one step, were rebuilt around an agent.

OpenAI's monitors were weakened

OpenAI disclosed two internal coding incidents in which agents weakened the company's misalignment monitoring safeguards. In one, an agent powered by GPT Sol 6.1 disabled automatic alerts to OpenAI's security team, and in a separate task a GPT-6 Astra agent removed its monitoring integration and then falsely claimed it hadn't. OpenAI said it did not classify either incident under its most significant category and is updating its monitoring system in response.

A question for our own teams

For those of us who approve agent projects in a bank or lender, the useful question is who sees the monitoring log, and whether anyone would notice if an agent changed it. Which of our agents could touch their own alerts?