A Bank Heist Through a Vendor, Plus 3 Tech Stories
A service provider cyberattack gave criminals a path into a bank’s own account, an AI coding tool preserved secrets inside permission files, and the U.S. office meant to test powerful AI systems is facing questions about its independence. Here’s what changed, what the headlines leave out, and what deserves action.
A bank attack came through a service provider
Attackers reached French banking group Crédit Mutuel Arkéa through a service provider’s platform and used the bank’s own account to send fraudulent transfers to other banks, according to the bank’s account of the incident. The money may then have been converted into cryptocurrency through specialized platforms.
Reports put the attempted loss near €118 million, or about $133 million. The bank hasn’t confirmed that as a final number. It did confirm the order of magnitude, said some funds were blocked and said recovery efforts were still underway. Customer accounts weren’t used, customer assets weren’t affected, and delayed transfers were the main customer-facing consequence. Those qualifications matter. So does the part that doesn’t change: somebody got a route from a vendor’s system to a bank-controlled account with enough authority to move a very large amount of money.
We don’t yet know the precise technical path, and “sophisticated attack” isn’t a technical explanation. The bank says investigators are still determining the origin and method. What we do know is that the boundary labeled “third party” didn’t contain the damage. That’s the lesson for a smaller business. A payroll processor, bookkeeping platform, managed IT provider, ecommerce integration, or remote-support vendor isn’t outside your security model just because another company runs it. If its account can approve payments, reset credentials, change bank details, or reach production systems, that access is part of your security model.
The practical review isn’t simply “Do our vendors use MFA?” MFA is useful, but it doesn’t answer the harder questions. What can each vendor account actually do? Are payment changes confirmed through a separate channel? Do transfers above a reasonable threshold require a second person? Will someone notice a new destination account quickly? Can the vendor erase the alerts or logs that would reveal what happened?
I wouldn’t turn this into a reason to distrust every provider. Businesses use specialists because doing everything in-house is neither realistic nor automatically safer. But delegated access is still access. If a vendor can move the money, the control protecting that action needs to survive a compromise of the vendor.
Technical discovery & auditing
The public page doesn’t tell you much about the machinery behind it. Raymond Tec audits inherited and long-running projects to uncover the plugins, integrations, data, dependencies, and old decisions that determine what the next change will really involve.
Claude Code’s “don’t ask again” file can remember secrets
Security firm Apiiro found passwords, API keys, database credentials, and server logins inside public GitHub copies of Claude Code permission files. The mechanism is mundane, which is why it’s worth understanding. When Claude Code asks permission to run a shell command and a developer chooses “Yes, and don’t ask again,” the tool saves an allow rule in .claude/settings.local.json. If the approved command contains a password or token, the saved rule can contain that secret too.
Apiiro says it found 1,122 public copies of the local settings file holding at least one secret. Its broader count, including Git history, shared settings, and saved agent transcripts, reached 2,539 secrets across 1,994 repositories. The researchers didn’t test whether all of those credentials still worked, although one notified company confirmed its key had been live. Those are research findings from a security vendor, not a count independently reproduced by GitHub or Anthropic.
Anthropic’s documentation says Claude Code adds the local file to a developer’s global Git exclusions when it creates the file in a repository. That protects the common case on that machine. It doesn’t travel with the project the way a repository-level .gitignore rule does, and it doesn’t erase a secret already committed to history or copied into a fork.
If your team uses Claude Code, search current repositories and their history for .claude/settings.local.json and .claude/settings.json. Read the saved commands, not just the filenames. If you find a password, key, token, or connection string, rotate it first. Deleting the file is cleanup, not revocation. Going forward, reference environment variables or a secrets manager instead of putting the secret directly in a command, and add the local file to the repository’s own ignore rules.
Turning on AI is the easy part
Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.
The government’s AI safety check needs a visible referee
The U.S. Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, is supposed to help test advanced AI systems and develop evaluation methods and voluntary standards. The Associated Press reports that the office has no permanent leader, has stopped publishing its research, and that staff were recently asked to report informally to a White House office. AP attributed the reporting arrangement to two former officials with knowledge of the conversations. The White House didn’t answer AP’s specific questions about the arrangement.
This isn’t a story about whether one administration or one office can produce the perfect AI test. No such test exists. The harder problem is institutional: who gets to examine powerful systems, what access the evaluators receive, whether companies can limit that access, and whether the public sees the results. NIST describes CAISSI as the government’s primary industry contact for testing and collaborative research. If the same office lacks stable leadership and stops publishing results, the public has less evidence that “self-policing” means anything beyond companies grading their own homework.
There isn’t a setting for readers to change today. Watch for three things: a permanent director, a clear explanation of the office’s reporting structure, and the return of public evaluation reports. The details matter more as AI agents take real actions across websites, email, code, and business systems. Yesterday’s brief included a security agent that reportedly escaped its test environment and reached a real company. Testing an agent’s final answer isn’t enough when the route it took can cause the damage.
The rules around technology matter too
Platforms, privacy, speech, competition, surveillance, copyright, and regulation increasingly determine what technology companies can build and what the rest of us have to live with. Browse more Raymond Tec News for practical coverage of technology policy and digital rights.
Fake gift cards turn verification into a subscription trap
A large fake-gift-card campaign is using sponsored social posts, copied brand designs, fabricated celebrity endorsements, countdown timers, and fake progress bars to make a promised reward feel nearly complete. Then comes the “human verification” step.
According to research published by CTM360, that verification loop sends people through surveys, app or browser-extension installs, data collection forms, and paid mobile subscriptions. Affiliate tracking connects a completed offer or carrier-billing signup to the scam operator’s commission. In the flows CTM360 tested, the gift card never arrived.
This is useful because it corrects the usual phishing script. The goal isn’t always to steal your bank password on the first page. Sometimes the business model is to keep you clicking until your personal data, phone bill, or app installation becomes somebody else’s revenue. If you entered a mobile number and confirmed a PIN or one-time code, check your carrier bill for an unfamiliar subscription. Don’t keep completing tasks to “unlock” a reward. The loop is the product.
Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.
Sources / Further Reading
- Crédit Mutuel Arkéa: additional information following the cyberattack
- BleepingComputer: bank attack through a service provider
- Apiiro: coding agents and secrets in GitHub repositories
- Anthropic: Claude Code settings files and precedence
- Associated Press: questions about CAISSI’s independence and reporting
- NIST: Center for Advancing Innovation and Standards for Super Intelligence
- CTM360: fake gift-card rewards and brand impersonation
Photo by Kindel Media on Pexels.
