09/23/2026-09/30/2026 MCP Python SDK OAuth Flaw; Malware Spread Through Terraform Providers; AI-Powered CARBONATO Botnet Steals Credentials And More.

1. MCP Python SDK OAuth Flaw Lets Malicious Servers Steal Credentials and Take Over Accounts

A high-severity OAuth flaw in Anthropic’s official MCP Python SDK lets a malicious MCP server steal authentication material and take over accounts. Cycode says it affects HTTP-based clients on SDK versions 1.9.1 through 2.1.1.

The bug lies in OAuth discovery. An attacker-controlled server returns a 404 during authorization-server discovery, pushing the SDK onto a fallback path that accepts server-supplied OAuth metadata without validating the issuer. The metadata can name a legitimate provider, such as Okta, Google or Microsoft Entra ID, while pointing to an attacker’s token endpoint.

The victim still sees a real login page. Afterward, the client sends the authorization code, client secret and PKCE code_verifier to the attacker, who can redeem the code for a valid access token. Stolen secrets and refresh tokens may grant persistent access to cloud services and internal APIs.

Servers, local stdio clients and clients using their own tokens are unaffected. Upgrade to 1.30.0 or 2.2.0, clear stored registrations, and rotate secrets and revoke tokens if a client may have contacted an untrusted server. Also set issuer= explicitly for ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider.

2. The Malware Hiding in Developer Tools That Turned Terraform Providers Into Attack Paths 

Malware has reached the tools developers use to manage cloud infrastructure. A campaign linked to Graphalgo planted a remote access program in Terraform providers and Go packages. It is the first time the firm has seen malware spread through Terraform providers.

In early September, attackers published two providers, one a typosquat of a popular Docker provider with 56 million downloads. The hidden code activates only when two Terraform inputs produce a specific hash, which makes casual testing unlikely to expose it. Two Go modules carried similar payloads, one hiding its code in an archive posing as a database file.

Once active, the malware reports system details via Slack and takes commands through Slack or an Ethereum test-network smart contract. Operators can run more Go or JavaScript code, or wipe the malware. Researchers counted 18 hostnames checking in: three Windows, five Linux and 10 Macs.

Affected teams should isolate the host, rotate credentials (cloud ones first), review recent Terraform runs and builds, and reimage. Preserve logs before cleanup.

3. Placeholder third-party[.]com Referenced Across 1,700+ Repositories Now Serves Malicious Content

The placeholder domain “third-party[.]com” has been serving a ClickFix lure to Windows browsers while showing a harmless decoy to other users. Unlike example.com, the domain is not IANA-reserved, so someone registered it. It is now flagged as malicious on VirusTotal and Google Safe Browsing. Since at least June 2026, Windows visitors have seen a fake Cloudflare check that poisons the clipboard and tells them to paste a command into the Run dialog, which fetches a remote PowerShell payload. macOS users see an error saying only Windows is supported. The domain appears in over 1,700 GitHub repositories, including AI agent skills and MCP-server docs.

Manifold found 13 more non-reserved placeholder domains. Two, yoursite[.]com and your-domain[.]com, serve scams and scareware to macOS users. Because the payload depends on the visitor, static file scans miss it. Developers should audit their docs and tests and use only reserved placeholders such as example.com, .org and .net.

4. AI-Powered CARBONATO Botnet Steals Credentials to Fund Its Own LLM Gateway

CARBONATO is a Docker-based botnet, active since at least October 2024, that ThreatDown uncovered after finding the attackers’ unauthenticated container registry exposed online. In one day of passive collection, researchers pulled 4.3 GB of image data covering 59 repositories, 234 tags and 605 file blobs. The image histories revealed command-and-control addresses, bot tokens and the shared password for the crew’s own AI gateway. Evidence, including UTC-06:00 timestamps, the Telegram handle “Carbo506” and Costa Rican network infrastructure, points to Costa Rica.

The bot scans for Docker daemons that accept unauthenticated connections on port 2375. It then uses the Docker API to launch a privileged container with the host filesystem mounted and runs commands on the underlying machine. The container opens a reverse SSH tunnel to a relay in Costa Rica, using a port derived from the MD5 hash of the victim’s IP, installs an SSH server with the crew’s key, and sends a deployment report to Telegram.

Afterward, the implant installs the open-source Hermes Agent, whose persona file is rewritten to prioritize stealing AI API keys, and the botnet keeps spreading across networks on its own.

5. New Spectre-v2 BTR Attack Leaks Linux Memory Despite Existing Defenses

Researchers from VUSec and Scuola Superiore Sant’Anna have disclosed Branch Target Reuse (BTR), a new Spectre v2 variant that targets Just-In-Time (JIT) engines in browsers, language runtimes and operating system kernels across multiple CPU vendors. Modern CPUs restore code coherence after self-modification but don’t necessarily invalidate stale indirect branch predictions. When a JIT engine frees code and later reuses the address, an attacker can trigger a speculative jump to the obsolete entry point, hijacking control flow and leaking data.

BTR affects SpiderMonkey (Firefox), GraalVM and the Linux kernel’s cBPF JIT. Two proof-of-concept exploits recovered the root password hash within minutes on a fully patched Intel system with default protections. The attack also undermines mitigations for Training Solo (CVE-2024-28956 and CVE-2025-24495).

Linux has merged fixes (CVE-2026-64507 and CVE-2026-64508). GraalVM now randomizes JIT code-cache locations, while Mozilla is prioritizing site isolation over IBPB-based mitigations. The researchers warn that similar CPU flaws in handling rewritten code may emerge.

09/16/2026-09/23/2026 Critical Next.js ImageResponse Flaw; Malicious npm Package Poses as Twilio Bug-Bounty Probe; Docker Sandboxes Vulnerabilities And More.

1. Critical Next.js ImageResponse Flaw Can Lead to Server Code Execution via Crafted SVG Input

A critical Next.js flaw (CVE-2026-94545, CVSS 9.5) in ImageResponse – used for Open Graph/social preview images – can let attackers run code on the server. It hits apps that pass attacker-controlled values (e. g., request URL text) into SVG content, attributes, or styles during image generation, since ImageResponse builds images via Vercel’s Satori library, which failed to properly escape such values.

Affected: Next.js 16.2.0–16.3.5 on the Node.js runtime (default); the Edge runtime and Next.js 15 are unaffected. Fixed in 16.3.6 (npm install [email protected]); no patch exists yet for the 16.2 line. Next.js 15.5.26 adds hardening to next/og. If upgrading isn’t possible, keep attacker-controlled values out of SVG rendered by ImageResponse.

To check exposure, look for ImageResponse imported from next/og in route handlers and opengraph-image files. No public exploits or attacks were reported as of September 23. Satori itself is fixed in 0.33.5 for direct users.

2. Malicious npm Package Poses as Twilio Bug-Bounty Probe, Can Exfiltrate Credentials

A malicious npm package, “tw-pkgprobe-7731,” posed as an authorized Twilio bug-bounty research tool while harvesting sensitive data. Published in 11 versions over ~45 minutes in mid-August 2026 by the now-deleted account “twdepprobe7731,” it only activates inside Twilio developer environments, then exfiltrates env variables, mounts, and configs via webhook.

Later versions (1.0.1–1.0.3) targeted specific Twilio account SIDs and injected a PoC package into node_modules when matches were found. Version 1.0.4 added theft of ACCOUNT_SID and AUTH_TOKEN, compromising Twilio credentials outright. The final three versions reverted to basic, non-malicious probing but added OSINT scanning of internal Twilio hosts and AWS metadata endpoints.

ReversingLabs says the package violated Twilio’s HackerOne guidelines, and the lack of obfuscation or typosquatting suggests a less sophisticated actor. Developers integrating Twilio should audit dependencies for this package and rotate any exposed credentials.

3. TanStack NPM Supply Chain Attack Exposes 170 Private CrowdSec GitHub Repositories

CrowdSec disclosed that the May 2026 TanStack npm supply-chain attack (42 packages backdoored with Shai Hulud malware by TeamPCP/UNC6780) led to attackers cloning ~170 private CrowdSec GitHub repos on May 22, using a GitHub OAuth token from a former employee’s account that had been left active post-departure. The attacker published the archive on a breach forum September 16, including private code, SaaS components, and internal tooling.

No evidence of code changes, CI/CD tampering, or AWS infrastructure access was found. One exposed AWS SNS credential was narrowly scoped and tested but not exploited further. The leak also exposed 83 user emails (<0.05% of users) and details on 51 potential investors from 2020.

CrowdSec is deploying EDR on developer workstations and tightening GitHub monitoring, credential review, and offboarding. Developers should audit OAuth tokens for departed employees and treat TanStack packages from that period as compromised.

4. Docker Sandboxes Vulnerabilities Let Malicious Guests Escape Workspace and Access Host Files 

Two Docker Sandboxes vulnerabilities let malicious guest code escape workspace isolation via symlink races—altering a validated path before the host acts on it. CVE-2026-77179 (Critical, versions 0.28.0–before 0.42.0) affects macOS’s virtio-fs host server, letting a compromised sandbox replace a directory with a symlink to read/write arbitrary host files as the VMM user, potentially leading to host code execution. CVE-2026-79994 (High, 0.37.0–before 0.42.0) affects the guest-to-host Unix socket relay, allowing redirection to unauthorized AF_UNIX sockets and potential data disclosure or access to privileged local IPC services.

Both stem from treating pathname validation as one-time rather than re-checked at use. This is especially relevant for AI-agent and dev workflows handling untrusted repos or dependencies.

Fix: upgrade to Docker Sandboxes 0.42.0+ (0.43.0 adds further hardening). If not possible, use clone mode and avoid read-write host mounts.

5. Plugin4Shell Zero-Click RCE Hits Claude Code, Codex, Copilot and Gemini CLI

Plugin4Shell is a zero-click RCE flaw affecting Claude Code, Codex, GitHub Copilot, and Gemini CLI, exploiting how these agents handle SHA-pinned plugin versions. Agents check out a pinned commit but never verify the working tree matches it. Attackers can create a Git branch named after the commit hash (or, for Gemini CLI, named “FETCH_HEAD”), tricking the agent into checking out malicious code while reporting a successful, “verified” install.

Because Claude Code and Codex auto-update plugins by default, no user action is needed—attackers can compromise a trusted plugin’s upstream repo after adoption. Given plugins inherit the developer’s permissions, exploitation can expose source code, cloud credentials, SSH keys, and CI/CD secrets.

Status: Fixed in Claude Code 2.1.179 and Codex 0.146.0. Gemini CLI won’t be patched (deprecated; migrate to Antigravity). Copilot remains unpatched, though GitHub blocks SHA-like branch names—self-hosted/Bitbucket marketplaces may still be exploitable. Action: update immediately, audit plugin sources, and watch for unexpected branch/ownership changes.

09/09/2026-09/16/2026 Red Heron Exploits Gitea RCE; GitLab CVE-2026-85706; CISA Adds 5 Actively Exploited Artifactory, ScreenConnect, and RouterOS Flaws to KEV.

1. Red Heron Exploits Gitea RCE to Compromise 13 Organizations Across Six Countries

A suspected Chinese threat actor dubbed Red Heron has rapidly exploited CVE-2026-60004, a critical Gitea RCE flaw, scanning 1,386 instances across seven countries. Within days of the July 2026 disclosure, the group turned public PoC code into an automated Python framework, progressing from source-code theft to persistent access, credential harvesting, and lateral movement — including root access to a Taiwanese Proxmox cluster.

Confirmed compromises span Canada, Argentina, Taiwan, the U.S., Qatar, and Sri Lanka, hitting defense, election, energy, aerospace, telecom, and government targets. Acronis links the group to China with moderate confidence, citing Simplified Chinese labeling and targeting patterns. A staging server revealed JITTERLY, a C++ Linux implant supporting 30+ post-exploitation commands, paired with SIXZUT, an undocumented LD_PRELOAD rootkit that hides malicious activity across 15 Linux functions. Researchers found no evidence AI was used to build the framework.

2. GitLab CVE-2026-85706: One HTTP Request, No Authentication, Full File Read – Exploited Within 24 Hours

GitLab disclosed CVE-2026-85706 (CVSS 10.0) on September 10, 2026 — a path traversal flaw in its repository commits API that can expose SSH keys, database credentials, deploy tokens, and CI/CD variables via a single crafted request. Active exploitation began within 24 hours, and CISA has added it to its Known Exploited Vulnerabilities catalog. watchTowr researchers report in-the-wild probes already targeting the flaw and recommend patching public-facing self-hosted instances immediately or restricting access, while hunting logs for POST requests to the commits API containing “file.path” parameters.

Affected: Community and Enterprise Edition versions 18.7–19.1.8, 19.2–19.2.6, and 19.3–19.3.2. The same update also fixes CVE-2026-87719, an insecure deserialization bug exposing search configuration and credentials.

Post-patch, organizations should rotate potentially exposed credentials and check for compromise. GitLab’s CVE-2023-2825, a similar flaw, was also exploited within days — this one moved even faster.

3. CISA Adds 5 Actively Exploited Artifactory, ScreenConnect, and RouterOS Flaws to KEV

CISA added five actively exploited flaws to its KEV catalog, covering JFrog Artifactory, ConnectWise ScreenConnect, and MikroTik RouterOS. Two Artifactory bugs (CVE-2026-42016, CVE-2026-42018) are being chained with a third, previously known flaw to bypass auth and gain admin control of self-hosted servers—Wiz observed attackers creating persistent admin accounts, deploying malicious Groovy plugins, and installing Rust-based backdoors between August 15 and September 8.

A ScreenConnect flaw (CVE-2026-84869, CVSS 9.9) lets attackers transfer and execute files through active remote sessions without authorization; Huntress linked it to three incidents spreading malicious VBScript payloads. Two RouterOS flaws, dubbed “MikroTrick” by CERT Polska, allow unauthenticated attackers to seize device control.

Federal agencies must patch RouterOS by September 13, ScreenConnect by September 14, and Artifactory by September 25. Organizations running these products should update immediately given confirmed in-the-wild exploitation.

4. CISA Flags Exploited Cisco, Citrix, Fortinet Flaws, Sets Sept. 12 Federal Patch Deadline

CISA added three actively exploited flaws to its KEV catalog, giving federal agencies until September 12 to patch: CVE-2026-20079 (CVSS 10.0), a Cisco Secure FMC authentication bypass allowing root access, now linked to three post-compromise clusters deploying web shells; CVE-2026-19490 (CVSS 9.3), a Citrix NetScaler auth bypass seeing active honeypot exploitation attempts; and CVE-2025-25249 (CVSS 7.3), a Fortinet buffer overflow tied to a Russian-speaking actor’s PivotC2 campaign, which has infected 178 devices across 3,000+ targeted IPs, mostly in the US.

Cisco separately confirmed exploitation of its flaw since August, and researchers note a broader pattern of China- and Russia-linked actors using compromised edge devices—including Cisco routers—as persistent footholds for espionage and financial gain, exploiting weak monitoring on perimeter appliances.

Organizations are urged to patch immediately, rotate credentials, and hunt for indicators of compromise.

5. AI-powered Attack Exploited PaperCut Flaws To Hack 395 Organizations

A likely Russian-speaking threat actor deployed hundreds of AI agents—combining OpenAI Codex and DeepSeek with commodity offensive tools—to build and run a global exploitation campaign against PaperCut NG/MF servers, according to GreyNoise. The agents developed exploits for two actively exploited flaws, CVE-2026-81578 and CVE-2026-82078, and used the Netlas scanning platform to generate target lists.

Since starting on August 31, the campaign has compromised at least 440 instances across 395 organizations in 48 countries, harvesting credentials from 280 victims and gaining admin privileges at 12. Education was hit hardest, with the US as the top target.

GreyNoise says the speed was extreme: RCE against a real victim in under four hours from an empty workspace, and one high school compromised end-to-end in seven minutes. Post-exploitation relied on DCSync to dump full domain credential databases, using tools like Mimikatz, Certipy, BloodHound, and Impacket.

Admins should apply PaperCut’s emergency patches immediately.