On September 3, 2020, a risk analyst at Credit Suisse escalated his worries about one of the bank’s hedge fund clients, Archegos Capital Management, to his supervisor. He sat in the bank’s independent credit risk team.
“Where am I going with this?” he concluded. “All of the people that I would trust to have a backbone and push back on a coverage person asking for zero margin on a heaping pile are gone.” The team he meant was the business’s own risk unit, which watched Archegos day to day and set its margins. It was now “run by a salesperson learning the role,” he went on. “I don’t consider PS Risk the best first line of defense function anymore.”
The analyst and his supervisor agreed that the business’s risk unit wasn’t taking sufficiently aggressive steps to manage the growing risk. When they followed up with the business, according to the investigation the bank’s board later commissioned, they were told progress was being made.
About six months later, Archegos told the bank it could no longer meet its margin calls. Credit Suisse lost about $5.5 billion. In 2020, Archegos had brought it about $16 million in revenue. The case has the rest.
Three roles
The analyst was using the language of a governance model that, according to the Institute of Internal Auditors, applies to all organizations. It used to be called the Three Lines of Defense. In 2020 the IIA updated it and renamed it the Three Lines Model, and it refreshed the model again in 2026.
The first line is the people who deliver the product or service. They take the risks, and in the words of the IIA’s 2020 paper (updated in 2024), they “remain responsible for managing risk,” whatever else is set up around them. The second line helps them manage it, with “monitoring, advice, guidance, testing, analyzing, and reporting,” and provides what the IIA calls “support and challenge.” The third line is internal audit, whose “defining characteristic” is “independence from management,” and which answers to the board.
The numbers don’t mean the lines take turns, the IIA adds. “All roles operate concurrently.” And it is frank about one weakness: because the second line is part of management, it is “never fully independent from management, regardless of reporting lines.”
Every line had a warning
At Credit Suisse, all three lines were in place, and each had a warning in front of it.
The first line had an in-business risk unit whose job was to set margins and watch client exposures. The second line, an independent credit risk team, set limits for Archegos and saw one of them breached virtually every week from mid-July 2020. The third line, internal audit, was asked after an earlier client default to audit how that credit risk team oversaw trading clients. In October 2020 it gave the team’s oversight a C-minus, finding 180 open limit breaches caused by new trading, on average 47 days old, and 824 more caused by market moves, on average 100 days old. The team was told to fix this. It hadn’t, by the time Archegos failed.
“The Archegos risks were identified and were conspicuous,” the board’s investigators wrote. What failed was the next step. When Archegos breached a severe stress test, the business argued for a milder one, and risk agreed. Risk, the investigators wrote, “adopted the business’s justifications for Archegos’s breaches and margin accommodations.” Switzerland’s regulator, FINMA, later put it more bluntly: “exceeded limits were simply increased repeatedly.”
The lines had also thinned. Staff told the investigators that both the first and second lines had been short of people and experience. When the head of the in-business risk unit died unexpectedly in 2020, he was replaced by a long-time sales and marketing executive with no in-business risk experience.
Three lines are not three filters that catch whatever the one before missed. A second line that takes on the first line’s reasons isn’t a second view anymore.
The football club
Imagine a youth football club run by volunteers. The coach buys kit and books pitches. The treasurer checks receipts against the budget. Once a year an outside examiner looks over the accounts. That is three lines in miniature. Now imagine the treasurer has known the coach for years and signs off each overspend as a one-off. The examiner might flag it, and the committee might mean to deal with it at the next meeting. The money might run out before then.
A checker that doesn’t hear the excuses
Anthropic faced a version of this when it let its coding agent, Claude Code, act without asking a person to approve every step. In its auto mode, the agent does the work, and a separate classifier checks each risky action, such as a shell command or a change outside the project, before it runs.
One design choice addresses the kind of failure seen at Credit Suisse. The classifier sees the user’s messages and the action itself, but not the agent’s own explanations. “We strip assistant text so the agent can’t talk the classifier into making a bad call,” Anthropic wrote. “We want it to judge what the agent did, not what the agent said.” If the agent keeps getting blocked, three times in a row or twenty times in a session, the agent is stopped and the decision goes to a person. When it runs unattended, with no person to ask, the process is simply ended.
Anthropic is open about the limits. On a set of real cases where the agent overstepped, the checker let 17 percent through, and Anthropic says it is “not a drop-in replacement for careful human review on high-stakes infrastructure.”
The analyst’s warning named the real problem. Credit Suisse still had a first line of defense on its organization chart. What it had lost, in his view, were the people with a backbone.