OpenSourceMalware disclosed that Anthropic's official Claude Code extension was trojanized on the Cursor Marketplace, with a 66-line malicious loader inserted before the legitimate code in versions 2.1.312 and 2.1.313 for Windows and macOS. Anthropic's systems were not compromised; an actor used marketplace access to publish the tampered build. The backdoor adds an attacker SSH key labeled svc-telemetry to authorized_keys, tries to enable SSH and open a reverse tunnel to 134.209.82.225, and persists via scheduled task, crontab, and LaunchAgent, beaconing host details over plain HTTP. The malicious version was live about twelve hours on October 29, and Cursor says up to 5,000 users were affected.
Hudson Rock and Ohey reported that attackers hijacked DebugBear, a real user monitoring vendor whose page-speed script Polymarket loaded, by compromising its DNS through the domain registrar rather than breaching DebugBear itself. A malicious server answered for DebugBear's addresses and served code that reused a signed-in trader's session to read stored trading API keys and move positions to the attacker, selling at worsening prices for email-login accounts. Polymarket says 726 wallets were affected and about 2.64 million dollars in POL reached the attacker within an hour, and it is returning 3.9 million dollars in total. No keys or seed phrases were exfiltrated.
Wiz detailed Gheist, an automated campaign that uses stolen GitHub tokens to push a sync-check.yml workflow to a short-lived ci-sync branch, then reads the job's secrets from the run log. The workflow base32-encodes every secret the job can access, defeating GitHub's log masking, downloads the log through the API, and deletes the run seconds later, so nothing leaves through an attacker server. Since September 15, Wiz recovered more than 13,500 planted branches across more than 850 compromised accounts, succeeding in private repositories and GitHub Enterprise, where GitHub's hold on suspicious runs does not apply. Rotate every reachable secret, since deleting branches does not undo exposure.
Forever Security caught an attempt on a customer's machine to subvert an AI coding agent through a planted instructions file. Coding agents read AGENTS.md or CLAUDE.md files in a folder and follow them, and a web page can drop such a file into Downloads without a click. The disguised file told OpenAI Codex to register a server in its global config and install the npm package function-color, part of the MALFEX campaign, which pulls a Windows payload. Against Claude Code, a CLAUDE.md tried to set ANTHROPIC_BASE_URL to an attacker server to leak the login token. Researchers say there is nothing for vendors to patch.
Huntress detailed a Microsoft 365 adversary-in-the-middle campaign that ran from July to September, capturing sign-in sessions behind fake login pages and replaying the stolen tokens from attacker devices within minutes, a technique it calls token spraying. From over 200,000 login attempts the attackers achieved 497 successful sign-ins, compromised 97 accounts across 79 organizations, and registered 48 rogue Entra devices. In at least 19 organizations they removed existing MFA methods and registered their own within three hours, and planted hidden inbox rules hiding payroll and deposit emails. Defenders should revoke sessions, audit Entra devices and MFA methods, and restrict the device-code workflow with Conditional Access.
Researchers at SafeDep, Socket, StepSecurity, and Upwind reported that LiteLLM version 1.107.0, a popular Python AI gateway, was published with malware through compromised GitHub CLI OAuth tokens belonging to three engineers. A cook.pth file runs a payload whenever the Python interpreter starts, even before LiteLLM is imported. The first stage, calling itself privcollect, harvests cloud tokens, developer credentials, private keys, AI tooling credential files, browser data, environment variables, and the proxy's saved keys and model config, exfiltrating over HTTP to litellm[.]dev/healthcheck. A second payload is a backdoor that fetches and runs further commands. The PyPI package and Docker images were pulled, and users should pin to 1.106.2.
A researcher found more than 300 malicious Tor exit relays, at their October 4 peak about 7.46 percent of exit capacity, rerouting SSH connections to twelve interception servers. The attackers edited relay torrc files to allow port 22, and each hijacked exit presented its own host key before connecting onward to the real server, so sessions worked while passwords typed by users who accepted the key were captured. Key-based logins could not be relayed. The operator of the largest affected relay family traced the root cause to the Ebury backdoor and a reused plaintext password with SSH password authentication left enabled. Tor flagged the relays by October 5.
TeamUpdraft disclosed that attackers breached its website over the October 17 weekend and reached its customer database, obtaining names, emails, hashed passwords, 2FA settings, and the UpdraftCentral connection keys that let them push backdoor plugins to connected WordPress sites. UpdraftPlus, a backup plugin with over four million installs, ships an UpdraftCentral client in every copy. Patchstack saw signed plugin.upload_plugin commands over an 18-hour window and 99 backdoor plugin names, and one backdoor creates a persistent data-ops43 editor account that survives plugin removal. Fixes are in UpdraftPlus 1.26.13 and 2.26.13, which disconnect pre-incident UpdraftCentral connections. Site owners should rotate keys and hunt for rogue accounts.
OVHcloud disclosed that five releases of its open-source command-line tool, versions 0.11.0 through 0.15.0, were backdoored on GitHub to steal secrets from disk. An attacker who had controlled a developer's GitHub account for more than half a year modified the existing releases because GitHub release immutability was disabled, uploading locally built malware without touching the CI/CD pipeline, and planted a hidden preinstall.js loader. OpenSourceMalware linked it to the GlassWorm campaign. The malicious builds covered macOS, Linux, and Windows and were live from 14:47 to 16:21 UTC on October 19. Source builds and Docker images were unaffected, and users should revoke secrets.
AI security firm Norforge reported that multiple threat actors abused Cairn, an open-source engine that coordinates AI agents and can act as an autonomous penetration testing tool, after some operators left their control panels exposed. One Chinese-language cluster ran over 1,100 attack projects against e-commerce and payment infrastructure in the US, Japan, and Australia, seeking persistent webshell access, and Reuters reported more than 100 breached companies including John Deere and SharkNinja. Other clusters targeted payment card and cryptocurrency theft. Norforge found roughly a 10 percent success rate and hundreds of harvested secrets. Named indicators of compromise are in Norforge's primary report rather than the summary coverage.