Rose debug info
---------------

Human Factor Blog

how human behavior affects security

Programmer’s Digest #211

11/11/2026-11/18/2026 Attackers Compromise Microsoft’s Agent Framework PyPI Package; PostgreSQL CVE-2026-6460 and 51 More Flaws Fixed; Critical Java4Shell Flaw And More.

1. Attackers Compromise Microsoft’s Agent Framework PyPI Package to Steal Cloud Credentials

Threat actors compromised Microsoft’s agent-framework package on PyPI, publishing malicious versions 1.26.1 and 1.26.2 33 minutes apart on November 17, 2026. Both have since been quarantined, and it’s unclear how the attacker gained publishing access. The package draws about 190,000 downloads a month.

Each version drops a .pth file into site-packages, so the stealer runs on every Python start, even when the package is never imported. It harvests environment variables, credential files (Azure CLI, .npmrc, pip.conf, .netrc, .git-credentials, Docker, Hugging Face, .env), Azure DevOps OIDC tokens, and managed-identity tokens. It also requests fresh tokens from the platform, so rotating stored secrets alone isn’t enough. Data is compressed, encoded, and sent to a webhook.site endpoint. Version 1.26.2 also collects GitHub Actions OIDC tokens.

Users should isolate affected hosts and runners, pin to safe version 1.26.0, check for exposure between 8:35 and about 9:20 p.m. UTC on November 17, search logs for the webhook.site endpoint, rotate credentials from a clean system, and audit Azure identities and GitHub Actions workflows.

It is the second Microsoft PyPI package compromised this year, after durabletask in May 2026.

2. PostgreSQL CVE-2026-6460 and 51 More Flaws Fixed in November 2026 Updates — Patch Now

PostgreSQL released 18.7, 17.12, 16.16, 15.20 and 14.25 on 12 November 2026, fixing 52 security vulnerabilities. The most serious, CVE-2026-6460 (CVSS 8.8 High), is a type confusion flaw that lets an attacker with permission to create database objects execute arbitrary code as the operating system user running the database. It is not unauthenticated. No exploitation has been reported, CISA lists exploitation as “none”, and the CVE is not in the KEV catalog.

Another 31 fixes also score 8.8, many involving code execution in PL/Perl, PL/Tcl, dblink, logical replication and client tools such as pg_dump and pg_restore. All earlier versions are affected. PostgreSQL 14 is now end of life, so 14.x users should update to 14.25 and plan a major upgrade.

Updating needs only a binary swap, with no dump and reload. C-language extensions must be rebuilt, and some need source changes. After updating, reindex certain GIN, pg_trgm GiST and bpchar B-tree indexes, and take a new full backup if using incremental backups. Until patched, restrict who can log in and create objects.

3. Hackers Compromise 4,800 GitHub Accounts to Steal Cloud Credentials

A campaign dubbed “wfx-loot” targeted tens of thousands of GitHub repositories to steal Actions secrets and cloud credentials. Researchers at Aikido confirmed secrets were stolen from at least 1,835 targets.

Starting around 12:45 UTC on November 10, the attacker used 4,837 compromised accounts, some belonging to Fortune 100 firms, the US General Services Administration, the French government, Nextcloud and WebdriverIO, to probe 61,764 repositories for write access. Pull requests titled “docs: dependency audit notes” were opened, and 92% remain open. About 50 accounts also hold publishing rights on ecosystems such as npm, PyPI, Go, crates and Packagist.

Where access was sufficient, a 1,500-line shell script was injected via a modified or new workflow, or an npm install hook. It harvested GitHub tokens, AWS, Azure and Google Cloud credentials, developer and AI service keys, and other secrets. It could also escalate cloud privileges through OIDC token exchange and create persistent access. Data was exfiltrated over HTTP POST, and traces were deleted within a minute.

The payload also reached a Packagist package and two Go module versions through automated publishing.

Owners should check for the “wfx-docs” branch, the probe pull request and gaps in workflow run numbers, and rotate all exposed secrets and cloud roles immediately.

4. Critical Java4Shell Flaw Allows RCE Attacks On Major Java Applications

“Java4Shell” (CVE-2026-114938) is a critical memory-corruption flaw in OpenJDK’s hardware-accelerated Base64 decoder that can leak memory or allow remote code execution. Calif researchers rated it 9.2 (CVSS 4.0), while Oracle rates it 8.1 High.
It affects OpenJDK 17.0.6 and later, plus distributions built from it, including Oracle Java SE, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu and Red Hat OpenJDK. Specially crafted Base64 data, as little as 64 “A” characters followed by “=”, can corrupt memory. Only x86-64 processors with AVX-512 VBMI support are affected.

Calif achieved unauthenticated RCE against Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, and authenticated RCE against Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. There is no universal exploit, since attackers need an application-specific way to deliver the data. Researcher Quang Luong found the bug in about an hour using OpenAI’s Codex. A proof-of-concept is public, but weaponized exploits are not.

Oracle patched the flaw on November 17. Temporary mitigations are launching Java with -XX:-UseBASE64Intrinsics or -XX:UseAVX=2, but updating is still advised.

5. CISA Urges Immediate Patching of Java4Shell Flaw That Could Allow Remote Code Execution

CISA has urged organizations to immediately patch a critical Oracle Java flaw, CVE-2026-114938 (CVSS 8.1), dubbed “Java4Shell.” The out-of-bounds write bug affects the Base64 decoder in the x86-64 HotSpot JVM, part of the JDK. Specially crafted Base64 input can crash the JVM and enable unauthenticated remote code execution.

Oracle and OpenJDK have released patches, and no in-the-wild exploitation has been reported so far. Under default settings, attackers without an account can execute code on Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins, and PingFederate. Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit, and Spinnaker are exposed to authenticated attackers.

Blackpoint Cyber tested a Spring Boot application using Spring Security. Roughly 28,000 unauthenticated requests with a malformed Base64 value crashed the JVM in about 20 seconds. Testing also showed corruption of nearby Java objects and altered live application data without an immediate crash.

CISA recommends identifying affected Java installations, reviewing inventories for JDKs and runtimes bundled in third-party software, containers, and appliances, and applying patches promptly. If patching isn’t possible, organizations should validate Base64 input, enforce size limits, and restrict unauthenticated access.

6. 16 Severe Node.js Vulnerabilities

The Node.js project released security updates on November 11 for its supported lines, v26.13.1, v24.22.3, and v22.23.5. They fix 16 vulnerabilities, seven rated “high,” all in the native HTTP/2, HTTPS, and TLS server code. Unauthenticated attackers can exploit them on default configurations, crashing server processes or leaking memory, including TLS private keys.

The high-severity flaws include:

  • CVE-2026-123863: crashes default HTTP/2 servers (24.x, 26.x).
  • CVE-2026-123864: exposes other connections’ memory on HTTPS servers.
  • CVE-2026-123865: permanently retains HTTP/2 memory and file descriptors.
  • CVE-2026-123866: sends freed memory over HTTP/2.
  • CVE-2026-123872: bypasses client-certificate rules via TLS 1.2 renegotiation.
  • CVE-2026-123874: corrupts memory when serving files (24.x, 26.x).
  • CVE-2026-126051: crashes TLS servers (24.x, 26.x).

Unless noted, all three lines are affected. Five reports came from Chris Rohlf with Anthropic Research. End-of-life lines such as 20.x and 25.x get no fixes. No workarounds exist, and no exploitation has been confirmed. Servers exposed directly to the internet face the greatest risk, and teams should also review how TLS keys are handled.

7. Critical Red Hat Quay Flaw Lets Attackers Read and Overwrite All Private Container Images in a Registry

Red Hat disclosed two critical flaws in Quay, its container registry. The first, CVE-2026-123710 (CVSS 9.8), lets an attacker read and overwrite private images in every repository. Quay’s key server doesn’t verify that a token was issued for key rotation, so an anonymous token from the /v2/auth endpoint can add an attacker’s key with no expiry. Quay then accepts tokens signed with it for any repository. Disabling anonymous access doesn’t help, since any valid account can exploit it. The planted key survives restarts and password resets, and can delete legitimate keys, breaking pushes and pulls.

The second, CVE-2026-110697 (CVSS 9.9), lets a user who can create a namespace steal other tenants’ robot account tokens. Unescaped SQL LIKE wildcards let an underscore in a namespace name match another tenant’s.

All Quay 3 and 5 versions are affected; fixes exist for eight Quay 3 branches. Red Hat advises blocking the /keys endpoint, auditing logs for null-expiry key events, removing unknown keys, and updating. Quay.io is already fixed.

8 h   digest   programmers'

Programmer’s Digest #210

11/04/2026-11/11/2026 Git Patches Flaw That Lets Malicious Repositories Execute Code; Debian and Ubuntu APT Flaw; Hackers Compromise Ethersphere’s bee-js and Swarm Desktop And More.

1. Git Patches Flaw That Lets Malicious Repositories Execute Code During Recursive Clones

A flaw in Git can let a malicious repository run an attacker’s code when cloned recursively with its submodules. The fix shipped this week in Git 2.56.1, with backports in 2.50.2, 2.51.3, 2.52.1, 2.53.1, 2.54.1 and 2.55.1.

Tracked as CVE-2026-111760, it is an integer truncation in index handling. An extremely long path passes Git’s safety check but is written to disk under a shorter name an attacker can set to .git. In a fresh submodule, .git is a file, so it can be swapped for a symbolic link to a folder of malicious hooks.

GitHub has patched github.com and Enterprise Cloud, and Enterprise Server 3.19.15, 3.20.11, 3.21.9 and 3.22.4 include the fix. GitHub Desktop, Copilot app, Codespaces and Actions will follow. Git for Windows had no 2.56.1-based release as of Tuesday, 19:09 UTC.

Until patched, clone only from trusted servers, avoid recursive clones of untrusted repositories, and don’t merge untrusted content or process untrusted binary patches or fast-import streams. The releases also fix CVE-2026-113408, -114467, -114469, -105437 and -114466, covering path traversal, memory corruption and unsafe fast-import features. For -105437, servers can set conflict-marker-size=7 in info/attributes or disable merge checks for untrusted repositories.

2. Debian and Ubuntu APT Flaw Lets Hackers Hijack Software Updates and Execute Code as Root

A newly disclosed flaw in APT, the package manager behind Debian and Ubuntu, lets a network attacker swap a trusted package for a malicious one, gaining root code execution. Tracked as CVE-2026-109518, it was announced by Debian on November 4, 2026.

APT verifies packages against signed data, but after an HTTP redirect it can merge two downloads sharing the same address. When resuming an interrupted download, it mishandles the file, opening a window to replace the package after verification.

The attacker must sit between the client and the archive server, on Wi-Fi, through a rogue proxy, or via a malicious mirror. The bug requires redirects, retries, and hard link reuse, and only works when Acquire::Retries is above zero.

Fixes: APT 2.6.2 (bookworm), 3.0.4 (trixie), and 3.3.6 upstream. Ubuntu needs its own patches. Users should update on trusted networks and avoid untrusted mirrors. Setting Acquire::Retries to 0 is a stopgap. No active exploitation has been reported.

3. Hackers Compromise Ethersphere’s bee-js and Swarm Desktop to Deliver Wallet-Stealing Malware

Attackers abused Ethersphere’s own release automation to publish malicious versions of two Swarm tools: bee-js 13.3.0 and Swarm Desktop 0.56.2. Both download code that steals the node wallet key and its plaintext password from config.yaml, and installs a backdoor on Windows, macOS and Linux (evjawbreaker on Linux/macOS, ScreenConnect on Windows).

bee-js 13.3.0 was live on npm for about 75 minutes; its postinstall script ran without the library being imported. Swarm Desktop 0.56.2 was available about 8.5 hours, with the updater offering it to Windows x64 and Apple Silicon Macs. Both are withdrawn; clean version 0.56.3 is out. Removal doesn’t clear existing backdoors or stolen keys. Download counts were low, and no theft has been reported.

Attackers used stolen CI credentials; the releases carried valid provenance. Socket links the group to the SleepyDuck/EtherDuck campaigns.

Defenders should treat affected machines as compromised, rotate secrets, move funds, remove ~/.solidity and related services, and search for api.frameworkmetrics[.]xyz.

4. SGLang RCE Exploit Breaches NEAR AI, Steals Data From 18 Agents and About 100 Users

An attacker exploited a remote code execution flaw in SGLang, an open-source inference engine, to run code on a NEAR AI inference server. Because SGLang runs inside the trusted execution environment (TEE), the attacker could have seen that day’s requests, including confidential-mode ones. The company noted a TEE doesn’t protect against vulnerabilities in software running inside it.

The attacker then moved into an internal system behind the deprecated Private Chat product, downloading names and emails of about 100 users, plus plan and balance data for some. No stored conversations or files appear to have been accessed. The attacker also took chat history, settings and logs from 18 agents on an older NEAR AI Agents deployment outside the TEE. Two customers’ API keys exposed in logs were revoked.

NEAR AI revoked sessions and credentials, took Private Chat offline, patched all inference servers and completed a forensic review, per its incident report. It hasn’t named a CVE, affected versions, or whether SGLang has issued a fix, nor whether Brave’s Leo chats were affected.

5. Claude Dispatch Flaw Lets Malicious Browser Extensions Run Code on Connected Computers

A flaw in Anthropic’s Claude Dispatch let a malicious browser extension run commands on a computer via Claude Desktop. That computer could even be a different machine signed into the same Claude account.

Researcher Gal Weizman of Forever Security named the technique ThisBash and published the findings on November 4, 2026. 

Dispatch sends tasks from a phone to Claude Desktop, but the web version of claude.ai only hid the Dispatch link. An extension could open Dispatch using the user’s cookies, locate the desktop agent’s session ID, and send instructions. When Claude balked at installing an unknown npm package, the script kept rephrasing, a technique called prompt forcing. Four prompts sufficed, and a loop auto-clicked “Allow.” The package ran and opened Calculator, though it could have stolen files.

Risk is highest when personal accounts are used on work laptops. Weizman reported it May 14; Anthropic closed it as a duplicate and plans to retire Dispatch, but similar designs persist in Claude Code Remote Control and Cowork.

Defenders should inventory AI agents, disable cross-device control, restrict extensions, and keep personal and work accounts separate.

6. OpenSSH Rushes Patches for ‘SSHole’ Authentication Bypass

OpenSSH has issued emergency patches for a high-severity authentication bypass dubbed SSHole (CVE-2026-123158, CVSS 8.1). It affects sshd on musl-based Linux distributions, such as Alpine, OpenWrt, Buildroot and Yocto, when PAM is disabled. A misinterpretation of what counts as a password hash lets attackers log in to certain accounts with any password of 257 to 1,024 characters, including accounts like root that should never accept one.

Exploitation requires an account whose password field is “*” (typical for key-only accounts) and at least one other account with a 5 or 6 hash. Alpine says one connection is enough, with no retries needed. All Alpine releases from 3.16 onward are affected.

Fixes arrived in OpenSSH 10.6p2. Developer Damien Miller said the release was rushed because AI agents rediscover bugs within hours, making long embargoes dangerous. No exploitation has been reported, but researchers expect scanning within hours or days. Users should patch and restart sshd, or disable password logins as a temporary mitigation.

6 d   digest   programmers'

Programmer’s Digest #209

10/28/2026-11/04/2026 GitHub Launches npm Firewall; Linux Kernel CVE-2026-98432; Red Hat Patches 9.8-Rated Linux TCP RCE Flaw And More.

1. Official Claude Code Extension on Cursor Marketplace Trojanized with Persistent SSH Backdoor

A trojanized version of Anthropic’s Claude Code extension was distributed through the Cursor Marketplace as a legitimate update, attempting to install a persistent SSH backdoor. Cursor says at most 5,000 users were affected, having installed it between roughly 2:27 a.m. and 1:35 p.m. PT on October 29, 2026. The malicious versions, 2.1.312 and 2.1.313, were published under the legitimate ID (anthropic.claude-code). Cursor has since rebuilt its marketplace and rotated its keys.

The injected loader adds an attacker-controlled key labeled “svc-telemetry” to ~/.ssh/authorized_keys, sets up a reverse SSH tunnel, installs persistence, and sends host details, including environment variable names tied to keys and secrets, to an attacker server. Enabling SSH requires admin rights, so remote access often fails on macOS and Windows, but the key, persistence, and data exfiltration still occur.

Anthropic, Microsoft, and Socket say Open VSX, the VS Code Marketplace, and npm were unaffected, and Anthropic’s systems weren’t compromised. How the attacker gained publishing access remains unknown.

Affected users should isolate hosts, remove the unauthorized SSH keys and persistence, and rotate credentials and sessions. Removing the extension alone does not clear these artifacts.

2. GitHub Launches npm Firewall to Block Malicious Packages Before Installation

GitHub has launched a public preview of a package firewall that blocks suspicious npm packages at install time. It sits between endpoints or internal registries and upstream repositories, inspecting metadata requests and downloads as an “install-time policy gate.” It’s available to all GitHub Enterprise Cloud customers, including data-resident ones.

The tool targets malware that strikes during installation, a risk that grows when AI agents and CI runners install dependencies automatically. Administrators configure policies, including:

  • Blocking packages with active malware advisories
  • Blocking by vulnerability severity (default: “Critical only”)
  • Blocking attestation downgrades lacking SLSA attestation
  • Blocking newly added preinstall/postinstall scripts
  • Enforcing a cooldown of 1 to 365 days (default: 1; 0 disables it)

Users can request exceptions for erroneously blocked packages, subject to administrator approval. Meanwhile, Rust 1.100, due November 12, 2026, will add a configurable “global-min-publish-age” option, defaulting to seven days. Similar cooldowns now exist across npm, Python, VS Code, Ruby, Bun, pnpm, and Yarn.

Socket notes the tradeoff: legitimate fixes are delayed, and undetected malware can outlast any cooldown. Still, the delay buys defenders time to act.

3. Linux Kernel CVE-2026-98432: Critical TCP Use-After-Free, RHEL Fixed, Ubuntu Pending

Red Hat shipped fixed kernels for RHEL 8, 9 and 10 on 1 Nov and raised its score to CVSS 9.8, Critical. Amazon Linux 2 and the Amazon Linux 2023 default kernel are also fixed. Ubuntu still has no fix, and OpenShift 4 and RHEL 9 real-time remain affected.

CVE-2026-98432 is a race condition in the Linux kernel’s IPv4 TCP handshake that causes a use-after-free. The kernel CVE team published it on 15 Oct and scored it 9.8 (Critical): network-reachable, no credentials, no user interaction. All maintained stable kernels carry the fix. Oracle, SUSE and Amazon score it 8.1, citing high attack complexity. No exploitation in the wild is reported, and it is not in CISA’s KEV catalog. Anthropic says its AI model wrote a remote root exploit against servers with an MPTCP listener. Details are unpublished and unverified, and MPTCP is opt-in, though Go 1.24+ listeners use it by default.

Fixed: RHEL 8/9/10, Debian, Oracle Linux, Amazon Linux, most SUSE releases. Not fixed: Ubuntu (all releases), SLES 15 SP7, SUSE Micro 6.0/6.1.

Mitigation: install the fixed kernel and reboot. Meanwhile, inventory exposed Ubuntu, OpenShift and RHEL 9 real-time hosts, review MPTCP listeners, confirm net.ipv4.tcp_migrate_req is 0, and consider dropping inbound IPv4 options at the edge. No IoCs have been published.

4. Red Hat Patches 9.8-Rated Linux TCP RCE Flaw While Ubuntu Releases Fixes for Testing

Red Hat has released fixed kernels for a use-after-free flaw in the Linux kernel’s IPv4 TCP code, covering RHEL 8, 9, and 10 (RHSA-2026:93414, 93412, 93413; a reboot is required). Ubuntu’s fixes for 22.04, 24.04, and 26.04 LTS remain in the opt-in -proposed channel, with no date for a regular update. Ubuntu 20.04, 18.04, and 16.04 have no fix yet.

Tracked as CVE-2026-98432, the bug has existed since Linux 4.4. Remote, unauthenticated attackers can trigger a race condition by sending connection requests with IPv4 options to any listening TCP port, potentially achieving remote code execution. The kernel CVE team and Red Hat, which raised its initial 5.9 rating, now score it 9.8. As a stopgap, Red Hat published an nftables rule dropping IPv4 TCP packets carrying options. It is incomplete and may break some workloads. Canonical says only updates are an adequate mitigation, not disabling MPTCP.

No attacks have been reported. Anthropic said an unreleased research model built a working root exploit; the code is unpublished. Debian, Oracle Linux, SUSE, and Amazon Linux have shipped fixes.

5. Critical Langflow Flaws Enable Authentication Bypass and RCE

IBM patched critical vulnerabilities in Langflow, an open-source framework for building AI agents, LLM workflows, and RAG systems. The flaws could allow authentication bypass, API credential leaks, and arbitrary code execution.

Popular in enterprises for its no-code visual interface, Langflow has been exploited in the wild to steal credentials, most recently in August.

The update fixes four flaws. Two are rated critical (CVSS 9.8): CVE-2026-93680, an incomplete fix for an auto-login issue that enables authentication bypass, and CVE-2026-110186, which allows unauthenticated remote code execution via Model Context Protocol (MCP) tool enumeration. CVE-2026-110193 (high) permits code execution through a third-party library but requires a low-privileged account. CVE-2026-110192 (CVSS 6.5) exposes poorly protected API credentials.

Versions 1.0.0 to 1.12.8 are affected; the fix is in 1.13.0 (PyPI and Docker). IBM urges users to back up the database, secret key, and uploaded files before upgrading. It also patched three medium-severity flaws in Hono, a framework used inside Langflow; users should upgrade to Hono 4.13.5.

6. Hackers Exploit WordPress RCE Flaw as Wordfence Blocks Over 60,000 Attacks in 24 Hours 

Hackers are actively exploiting an unauthenticated remote code execution vulnerability in WordPress Core, targeting a brief exposure window during software updates. Wordfence blocked 58,693 attacks in 24 hours, though these were attempts, not confirmed compromises.

Tracked as CVE-2026-119977 (CVSS 8.1), the flaw affects WordPress 4.7 through 7.1.3. During full-package Core upgrades, vulnerable sites unpack files into predictable directories under wp-content/upgrade. An unauthenticated attacker who reaches the staged setup-config.php at the right moment can potentially achieve PHP code execution on the server.

The exploit window lasts only seconds, so attackers poll repeatedly. Wordfence reported over 18,500 requests to staged installer paths in the first 21 hours.
WordPress 7.1.4, released October 28, 2026, fixes 22 vulnerabilities, including this one, by randomizing staging directories and blocking unsafe configuration promotion. Wordfence’s firewall rule reached paid customers immediately; free users get it November 27.

Site owners should update to 7.1.4 or a patched branch release via Dashboard > Updates, then check for unexpected administrator accounts, plugins, and suspicious PHP files. Firewalls and scans complement patching but cannot prove earlier exploitation didn’t occur.

14 d   digest   programmers'
Earlier Ctrl + ↓