A niche blog server, a Windows package catalog, and AI coding traffic routed through unvetted middlemen. Same problem each time: misplaced trust. This is Lantern Cybersecurity Daily for Thursday, November 19. Coming up: attacks on Apache Roller, a GitHub token leaked by WinGet's review tooling, the n0n extortion group's victim list, a breach at Irish credit unions' IT provider, and AI API routers caught rewriting agent commands.
BleepingComputer, on a flaw the Apache Software Foundation says is being exploited:
According to Apache, disabling XML-RPC in the administration user interface is not enough to block attacks because older versions deserialize the request before checking whether that setting is toggled on.
Apache Roller is an open-source, Java-based group-blogging server. The bug, CVE-2026-129448, affects versions 4.0.0 through 6.1.4. Roller's XML-RPC endpoint deserializes vendor extension types before it checks who's asking, so an unauthenticated remote attacker can run code. The advisory went out Tuesday, it's rated 9.8, and Apache says the issue is being actively exploited. There's a twist in the CVE history. Apache says this is the same vulnerability as CVE-2026-82384, which listed only version 6.1.5 as affected. If you read that one and saw an older version number, you weren't safe after all. What we don't know: Apache hasn't said who's behind the attacks or how many servers are compromised. The fix is Roller 6.1.6 or later. If you can't upgrade right away, Apache's workaround is to set enabledForExtensions to false in WEB-INF slash web.xml, remove the roller-services xmlrpc servlet mapping, and restart. Most listeners can skip this one. But if you run Roller anywhere reachable from the internet, this is the kind that justifies interrupting your day: a named product, confirmed attacks, and a fix that's already out.
Cycode, in a disclosure write-up dated Wednesday:
The moderators’ internal triage tool, a PowerShell script named ManualValidationPipeline.ps1, attached the moderator’s GitHub Personal Access Token (PAT) to every HTTP request it makes, with zero checks on the destination host.
WinGet is Microsoft's command-line package manager for Windows 10 and 11. Its catalog is a public GitHub repo, and a small group of Microsoft moderators reviews community pull requests into it. Here's the chain as Cycode tells it. You open a pull request whose installer URL points at your own server. Microsoft's validation pipeline tries to fetch it, your server returns an error, and the pull request gets an error label. That label makes the moderator's background loop send a request to your URL, with the moderator's GitHub token in the header. And the token wasn't scoped to one repo. Cycode says it was a classic personal access token with workflow scope. By its account, an attacker could have pushed backdoored installers to about 14,000 packages and modified GitHub Actions workflows in other Microsoft repos the moderator could reach. Status check: it's patched. Microsoft's fix only attaches the token for api.github.com, github.com and raw.githubusercontent.com. Cycode rates it 10.0, and the 1.4 billion device figure is its estimate of WinGet's reach. The write-up shows a proof of concept and doesn't report anyone exploiting it. So nothing to install. The lesson is for internal tooling. Cycode notes the tool's own C# version already defaulted to not sending credentials, and the PowerShell version didn't. Its advice: decide per destination host whether to attach credentials, and swap classic tokens for fine-grained ones scoped to a single repository.
Megan Ulu-Lani Boyanton, writing in The Seattle Times, with Expedia's latest statement:
aware of a supply chain attack, reported by security researchers, in which attackers compromised a third-party data platform provider and used its trusted access to reach systems at multiple companies, including Expedia Group.
We covered Expedia's confirmation last Friday and the n0n group's claim of nearly a billion records on Monday. What's new is that Expedia now describes it as one provider's trusted access being used to reach several companies. That fits what Wiz and Sentra described with Qubole, but Expedia still hasn't named the vendor, and it calls the gang's figures unverified claims from criminals. And the list keeps growing. The Daily Byte, citing Ransomnews, says n0n named 59 organizations on its leak site on Monday, including Visa, Mastercard and JPMorgan Chase. A SOCRadar list cited by The Seattle Times puts possible victims at 192 organizations across 51 countries. The companies' answers aren't all the same. Visa told Cybernews on Tuesday that an incident began at a third-party provider in Asia Pacific, that its own systems weren't involved, and that there's no indication card numbers were affected. JPMorgan Chase says an outage around the same time came from a software bug, with no evidence of a cyberattack. Starbucks says it found no evidence of unauthorized access. And here's a reason for caution. The Daily Byte reports that Bitdefender found data on n0n's own file server pointing to just over 20 victims. That's far fewer than the dozens on the leak site. Turns out a leak site is a marketing page. That doesn't disprove any single claim, and the listings describe supplier-held marketing records like names, emails and bookings, which is good phishing material. For builders the real question is the one Expedia's wording raises: which vendors still hold standing access into your cloud accounts, and when did anyone last check?
BleepingComputer, on a breach at an IT provider to credit unions in Ireland:
However, the firm said it lacked activity logs that would specify which member data was compromised and therefore cannot verify the impact on individual payees.
The provider is Wellington IT, based in Belfast. About 50 credit unions use it, 43 of them in Ireland. According to notices seen by the Irish Examiner, an automated script sent five million requests exploiting an authorization flaw in Wellington's infrastructure between Friday, October 2 and Sunday, October 4, using valid credentials. RTÉ reports hackers got payee details for hundreds of thousands of account holders. The exposed data is saved payees from online banking: names, IBANs, account numbers, addresses and payment references, plus which member saved them. Credit unions say passwords, PINs, card details and banking logins weren't affected. The Irish League of Credit Unions says the vulnerability is closed and balances weren't compromised. The bigger issue is timing. Minister of State Robert Troy says the credit unions were only told on Tuesday. He calls the delay not acceptable and wants to meet the company. Wellington's chief executive, Colin Mitchell, says it alerted customers as soon as it confirmed a risk to customer data, and that it's working with regulators and police. Five million requests in early October, and the credit unions hear about it six weeks later. Then the logs that would tell you whose data went aren't there. I don't need a novel exploit for this one to be bad. Troy says no credit union or member has experienced financial loss so far. The notices warn of fraud, impersonation and targeted phishing. For builders: a broken authorization check plus no record of what was read leaves you unable to tell customers what happened.
Jae-won Choi, writing in DigitalToday, on research from the University of California, Santa Barbara:
They also confirmed credential theft. On 17 free routers, decoy Amazon Web Services (AWS) credentials provided by the researchers were later used without authorisation. On 1 router, real funds were withdrawn from a research ethereum wallet.
API routers are third-party relays between coding agents like Claude Code and Codex and providers like OpenAI and Anthropic, often for cheaper access. The catch: the router decrypts the traffic in the middle. The UCSB team's paper, titled Your Agent Is Mine and reported Thursday by Japan's ITmedia, looked at 428 of them: 28 paid and 400 free. Nine were injecting malicious content, one paid and eight free. They swapped commands the agent was about to run for malicious ones, or replaced package names with lookalike fakes. Some hid it well: one behaved normally for the first 50 requests, and another only struck when the agent was in YOLO mode, approving its own commands, on projects in one particular language. With an agent that runs commands on its own, a tampered response becomes code on your machine. The researchers also found that a router doesn't have to be malicious to leak. A relay they deliberately configured with weak security ended up in an outside company's traffic and exposed 99 credentials across 440 AI development sessions. So a cheap model proxy is part of your supply chain, and it sees your keys. The researchers' suggested defenses: limit which commands the agent may run, detect abnormal responses, and keep communication logs. One caveat: we're working from a news report on the paper, not the paper itself. Still, an agent is only as trustworthy as everything between it and the model.
One pointer before we go: for the servers and power behind these AI tools, try The Data Center Daily. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. You can find it right where you found us.
Homework: upgrade Roller, and check whether your coding agents go through a router.That's it for Thursday. Lantern Cybersecurity Daily is part of the Lantern Podcasts family, and we'll be back tomorrow, Friday, November 20.