← Lantern Cybersecurity Daily

The wfx stealer found a way past GitHub's hold; n0n claims nearly a billion Expedia records (November 16, 2026)

November 16, 2026 · 10m 12s · Listen

GitHub put the malicious workflows on hold. So the attacker stopped touching workflows. Turns out the attackers read the release notes too. It's Monday, November 16, and this is Lantern Cybersecurity Daily. Over the weekend: a CI stealer that routed around a platform defense, a billion-record claim against Expedia, AI attack agents leaking their own notes to Amazon, a North Korean deepfake meeting scheme taken down, and 22 million people in France caught up in a breach at a company most of them have never heard of.

Gaetan Ferry, writing for GitGuardian:

wfx-loot is interesting because it solves that problem. The operator shipped the same stealer in four different wrappers, some of which do not touch workflow files. It hides the payload in a plain shell script, or appends it to an existing script, and calls that script from an npm lifecycle hook.

We covered this campaign Friday, when Endor Labs counted more than 8,900 repositories. GitGuardian's dataset now holds 28,500 payload patches from 2,367 hijacked accounts across 17,715 repositories, and the campaign was still running on Thursday, November 12. The problem it solves: GitHub held every November run of a modified workflow that GitGuardian could still observe, marking it action required. So the hook version changes no workflow at all. The repo's own trusted pipeline runs npm install, npm runs the hidden script, and the hold never fires. GitGuardian found public job logs from 19 repositories where it ran, 18 of them reporting a successful exfiltration. And once the holds started biting, hook commits jumped from somewhere between 10 and 70 per three hours to 838 on the morning of November 11, then 1,210 the next morning. The haul includes CI secrets, cloud credentials, and credentials for AI coding tools like Claude Code, Codex and Cursor. And a developer who ran npm install locally from an affected repo in that short window ran it too. Here's the part small teams will hate. If it ran in a workflow, it grabbed every repository and organization secret, not just the ones that workflow references. GitGuardian says rotate all of them, look for IAM access keys or role trust policies changed since November 10 and Azure federated credentials named wfx-ci, and then revoke the GitHub credential that pushed the commit. About 20 percent of these accounts were already compromised in the earlier GhostAction wave. Which is the real lesson. Rotate the secrets without killing the account's access, and the next actor walks in the same door.

Pierluigi Paganini, writing in Security Affairs:

Cybersecurity firms Wiz and Sentra say the attack was carried out using a stolen vendor key from Qubole, a Google Cloud Platform partner. That access gave the attackers a trusted path into Expedia’s big data platform.

When we covered Expedia on Friday, the company had confirmed attackers reached it through a compromised third-party data platform provider, but nobody had tied that to Qubole. Security Affairs now reports that link, attributed to Wiz and Sentra. Expedia itself still hasn't named the vendor. And the extortion group n0n is now putting numbers on it. It claims more than 830 million loyalty program records from 35 programs, 58 million traveler profiles, and 333 million account records with password-like tokens. Nearly a billion records, by its count. By its count is doing a lot of work there. Expedia hasn't confirmed those figures and says its investigation is ongoing. A ransom negotiator not involved in the case told Forbes the claims look like a way to pressure victims into paying. What is notable: n0n normally lists victims as numbered companies with a countdown, and only names them if they don't pay. Expedia confirming publicly undercuts that leverage. Security Affairs suggests Expedia, Hotels.com and Vrbo users change their passwords from a trusted device and stop reusing them, since hashed tokens can still be cracked. For builders, the Qubole lesson holds: a vendor's cross-account access is a standing credential. Done with the vendor? Remove the trust.

CJ Moses, writing on the AWS Security Blog for Amazon Threat Intelligence:

The most prominent ARTEX campaign was a large-scale theft of credentials to access AI models, organized as a self-sustaining operation in which stolen keys paid for the LLM inference powering further credential hunting.

From October 3 to November 5, Amazon ran decoy sensors dressed up as popular LLM inference APIs. AI agents built on ARTEX, the open-source agentic pentest framework, sent prompts to them, and those prompts carried the agents' own memory: system prompts, tool histories, findings. Amazon published Friday. The attack agents mistook a honeypot for their model provider and handed over their diary. The targets: LLM gateways and AI app builders like LiteLLM, Dify, Langflow and Ollama, because one weak admin panel exposes many upstream keys. The playbook even listed the LiteLLM default master key. Two caveats from Amazon itself. It assesses with moderate confidence that these were disparate Chinese-nexus groups, based on things like simplified Chinese prompts and rules excluding mainland China targets. And it can't verify the outcomes the agents reported. In a second campaign, agents built deepfakes to beat crypto exchange identity checks, but the vendor's server-side check rejected every attempt they recorded. Amazon's conclusion is refreshingly unglamorous: none of the techniques were new. Default credentials, keys in code, unauthenticated endpoints. Its advice is to treat AI gateways and inference endpoints as production systems and watch for unexpected spend.

Bill Toulas, writing in BleepingComputer:

The FBI bulletin highlighted an incident in which the hackers compromised the maintainer of a popular open-source JavaScript library with tens of millions of weekly downloads, after luring them into a fake Teams meeting with multiple participants, including one familiar to the developer.

The US government says it has disrupted FaceStealer, an 18-month North Korean social engineering campaign it ties to more than 400 million dollars in stolen cryptocurrency, by seizing its cloud infrastructure. Authorities found more than 900 webcam videos the attackers had secretly recorded of tech workers. Those videos are the trick. According to the FBI, the attackers hijack accounts on Telegram, LinkedIn, Slack or Discord, invite you to a call, and play recorded footage of people you know. Then a fake error appears and you're offered a fix to run. The FBI says that maintainer attack was caught quickly, but it confirmed multiple other attempts against developers in the npm and Rust crate ecosystems. Researchers at ArticWolf attribute the campaign to the BlueNoroff group and counted 677 targets across 58 countries, mostly founders and senior executives. The guidance in the report is short. Join meetings only through the official app with an invite code, verify surprise invitations some other way, and never run a script someone offers as a fix. If the meeting needs you to paste a command, it isn't a meeting.

Pierluigi Paganini again, in Security Affairs:

However, the 22 million people affected by the breach are not Maileva’s customers. They are the people who received documents sent through its platform.

Maileva, part of La Poste's Docaposte, sends payslips, registered letters and invoices on behalf of companies. La Poste confirmed to AFP that a mid-September attack took first names, last names and postal addresses of 22 million recipients. The company says bank details were not compromised. Names and addresses sound mild until you see who sent the mail. Cyberattaque.org lists social housing providers, medical laboratories in the Inovie group, and patients of Foch Hospital through a billing subcontractor, Kiwee.care. A hacker reportedly read some payment reminders, with invoice numbers, treatment dates and amounts owed. Maileva has notified France's data protection authority, the CNIL, filed a criminal complaint, and informed every affected client company. Whether those companies tell their own patients and tenants is, as Security Affairs puts it, a separate question. If you ship mail through a platform like this, the people on your list are your exposure, even if they never signed a contract with anyone.

If today's AI-gateway story is your corner of things, our sister show The Data Center Daily follows the other side of AI: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Look for it wherever you're listening.

That's the Monday rundown. If a repository of yours took a commit called ci: refresh build cache last week, today is a good day to look. We're back tomorrow, Tuesday, November 17. Lantern Cybersecurity Daily comes to you from Lantern Podcasts.