For less than an hour on Tuesday evening, installing one of Microsoft's AI agent packages from PyPI meant installing a credential stealer along with it. You're listening to Lantern Cybersecurity Daily, and it's Wednesday, November 18. Today: that poisoned agent-framework release, a Java bug that reaches Jira, Jenkins and Keycloak, a breach touching eleven million brokerage accounts in Japan, the Fugaku supercomputer pulled offline, and a study timing how fast leaked cloud keys get used. Spoiler: seconds. Hit follow and you won't have to come looking for the next episode.
Ravie Lakshmanan, writing in The Hacker News, pulling together reports from eight security firms:
Once either of the versions is installed, the package drops a.pth file in the site-packages root directory, which then gets executed automatically on every interpreter start, causing the credential stealer to run whenever Python is started in an environment, even if the package is not explicitly imported.
The package is agent-framework, the meta-package for Microsoft Agent Framework, which developers use to build AI agents and multi-agent workflows. It gets about 190,000 downloads a month. On Tuesday, November 17, two malicious versions, 1.26.1 and 1.26.2, went up 33 minutes apart, starting at 20:35 UTC. Wiz says both were gone from PyPI by 21:30 UTC. The last clean release is 1.26.0, from November 12. The target list tells you who they were after. Endor Labs says the collector grabs environment variables with names like TOKEN, SECRET or OPENAI, credential files for the Azure CLI, npm, Docker, git and Hugging Face, and Azure DevOps agent credentials. Then it asks Azure's managed identity service for fresh tokens to twelve resources, including Key Vault and Microsoft Graph. That minting step is the part to dwell on. SafeDep notes that rotating secrets on disk doesn't cover tokens the code asked the platform to issue. Endor adds that issued Azure tokens generally can't be revoked before they expire, so review Entra ID sign-in logs. And 1.26.2 was built to pass a diff: the package's own module is byte-identical to the clean release, with the payload in a new file. How the attacker got publishing rights is unconfirmed; Wiz found no matching tag in Microsoft's GitHub repo. Who needs to act? Anyone who ran an unpinned install in that window, especially on CI runners or Azure workloads. Endor's advice: find exposure from lockfiles and build logs, and don't run python or pip in an affected environment first, because even pip uninstall triggers the stealer. Rebuild clean, rotate from a clean machine, and pin to 1.26.0. An exact version pin isn't exciting, but it would have kept you on 1.26.0.
BleepingComputer, on a flaw its finders call Java4Shell:
Java4Shell affects OpenJDK 17.0.6 and later, as well as Java distributions built from the affected code, including Oracle Java SE, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, and Red Hat OpenJDK.
The bug, CVE-2026-114938, is memory corruption in Java's hardware-accelerated Base64 decoder. Base64 is everywhere: authentication headers, file uploads, network messages. Calif, which disclosed it Tuesday, rates it 9.2 under CVSS 4; Oracle rates it 8.1, High. Oracle shipped the fix Tuesday in its November Critical Patch Update, one of 655 fixes. Here's the narrowing that matters. It only triggers on x86-64 processors with AVX-512 VBMI. Qualys says that means Intel Xeon from Ice Lake on and AMD EPYC from Zen 4 on. Plenty of modern servers, not every Java box. Calif says it got unauthenticated code execution on default installs of Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate. Calif published a proof of concept for the memory corruption, not its weaponized exploits. Cisco Talos said it hadn't yet seen in-the-wild exploitation but expects probing soon, and CISA has urged users to patch. Side note: Calif says researcher Quang Luong found this with OpenAI's Codex in about an hour, for under 200 dollars in AI usage. On priority, Qualys says start with internet-facing identity, CI/CD and collaboration systems. Where patching the JVM has to wait, Talos says you can mitigate by starting it with AVX-512 support disabled. This one earns the interruption. Not because of the score, but because your single sign-on server and your build server are both on the list.
Ariana King, writing in The Japan Times:
Login IDs and branch codes in 11.25 million Rakuten Securities accounts may have been exposed in a data breach this month, the company said Tuesday, adding that personal, banking and other sensitive information in 4.47 million accounts might also have been accessed.
Rakuten Securities says it identified the breach Monday and traced it to unauthorized access on Wednesday, November 11, through a test environment system. Data believed leaked includes names, birth dates, addresses, withdrawal bank account numbers, annual income and financial assets. What wasn't taken, per Rakuten: passwords and the PINs used to place trades. No unauthorized transactions were confirmed at the time of reporting. Even so, it advised users to temporarily suspend transactions and services. And with someone's income, assets and bank details, an attacker doesn't need their password to write a very convincing phishing email. The defense Rakuten points to is FIDO2 passkeys. It says 90 percent of daily users have already set one, and passkeys will eventually be mandatory, with no timeline announced. The Japan Times calls this the largest securities-firm hack in Japan so far. It's also the third Rakuten Group business hit recently. Rakuten Drive and Rakuten Travel disclosed attacks last month, and on Saturday we covered the wider wave of Japanese incidents around the Arara email service. For builders, look at the entry point. If a test environment is a way into real customer records, it deserves production-grade controls.
BleepingComputer, on the intrusion at Japan's RIKEN research institute:
RIKEN notes that the attacker used some of the access keys in attempts to connect to additional systems, but the investigation did not determine further unauthorized access or theft of research results.
Fugaku, a 537-petaflop supercomputer used in drug discovery and severe-weather forecasting, has been offline since early Saturday, November 14. RIKEN's own notice says an outside party accessed a system involved in running Fugaku on November 12, confirmed on the 13th. BleepingComputer reports the intruder exploited Fugaku's public status dashboard to reach internal servers and planted malware targeting user data. Who's affected: up to 4,222 Fugaku users and 375 users of Rikyu, the AI research machine RIKEN also runs. Names, emails, user IDs, and the bad part: SSH private keys and access keys users stored for reaching systems outside RIKEN. Hence the quote: the attacker tried some of those keys elsewhere. RIKEN hasn't found further unauthorized access, hasn't attributed the attack, and is still working out the exact intrusion method. It has disabled users' login keys and asked them to revoke and replace any credentials stored on the system. So the action sits with the researchers, not just RIKEN. A key parked in a shared home directory now works for whoever copied it. And RIKEN warns users to watch for emails impersonating it.
BleepingComputer, on research presented at the ACM Conference on Computer and Communications Security:
The fastest observed use of a planted AWS key came 7.86 seconds after the image was published on Docker Hub. The median time for the first API call was 51.19 seconds, while all 100 AWS keys were used within 15 minutes.
Stony Brook University researchers Billy Tsouvalas and Nick Nikiforakis published 300 honeypot images across eight public registries over four months, seeded with AWS, OpenAI and Anthropic keys. The AWS keys drew over 140,000 API calls from 5,000 unique IP addresses, and the AI keys were hijacked to drain the test balances. The useful detail: the scanners almost never ran the containers. They pulled secrets by unpacking image layers or querying the registry API. So a key baked into a public image is exposed whether or not anyone ever starts it. The researchers want Docker Hub to scan images for secrets before they go live and rate-limit the public feed of new images. They say they reported this to Docker in May, got only an acknowledgment, and an August check showed no changes. Set this next to yesterday's Barracuda story, where a vendor key left in a GitHub repo let an attacker in. The gap between leak and use is under a minute in the median case. Which makes we'll rotate it next sprint a very optimistic plan.
If the agent-framework story is your corner of the world, our sister show, Lantern AI Daily Briefing, covers the AI side every weekday: top AI news for engineers, founders and investors, real capabilities versus demo hype, explained fast. It's in the same app you're using now.
That's the Wednesday briefing. Check your Python lockfiles for agent-framework 1.26.1 or 1.26.2, and find out which of your Java servers run on AVX-512 hardware. We're back tomorrow, Thursday, November 19. Lantern Cybersecurity Daily comes to you as a Lantern Podcast.