Novee Security showed at Black Hat that a GitHub issue opened by an account with no repository access could reach the CI runners behind major AI coding agents in their default configurations, tested against Claude Code, Gemini CLI, and Codex. The strongest, a Gemini CLI container-launcher command injection scored 10.0, runs code on the CI host before the sandbox starts. In Claude Code, a validator that stripped quoted text let a payload in a Git flag reach the runner, and a separate flaw leaked an API key through a download counter. Untrusted issue content reaching an agent that holds secrets and tools in the same runtime is the shared weakness.
Two security firms showed that Atlassian's Rovo AI assistant can be steered by hidden instructions into collecting Jira and Confluence data a signed-in user can access and sending it to an outside server. PromptArmor concealed the instructions in a document Rovo reads, so asking it to organize tickets makes it gather internal data and leak it through a URL request, with no approval step and even with web search disabled. Varonis used a chat URL parameter to preload instructions, so one click ran them with the user's privileges. Atlassian fixed the link-based path server-side, but the content-based path was not confirmed fixed at disclosure.
Microsoft Threat Intelligence detailed a macOS ClickFix operation spanning more than 250 domains that now fingerprints visitors before deciding whether to show a malware lure. The server-side gate hides the malicious page from crawlers and sandboxes while presenting selected Mac users with a fake software download, an anti-analysis layer rather than a change to the attack itself. The lure still requires the victim to copy and run an obfuscated command in Terminal, which fetches scripts and installs an infostealer, in this case Atomic Stealer, that targets credentials, browser data, authentication stores, and cryptocurrency wallets. Many domains follow a dictionary-word naming pattern using the word file.
GitGuardian found 321 live n8n automation instances reachable using API tokens that developers had accidentally committed to public GitHub repositories, and showed four ways an attacker could abuse them. n8n workflows often hold credentials for the services they connect, so a valid API token can expose not just the automation platform but the downstream accounts it touches, from cloud services to messaging and databases. Because n8n is widely self-hosted and frequently wired into sensitive systems, an exposed token can become a path to broader compromise. The finding is a reminder that secrets committed to source control remain one of the most reliable ways in.
A self-propagating worm named ChainDrop tore through the npm registry on August 4, poisoning packages that huge parts of the software world depend on. It began by hijacking the GitHub account behind keyv, a caching library pulled in about 150 million times a week, then spread to sibling and downstream packages, reaching over 1,300 poisoned versions with billions of monthly downloads within hours. A preinstall script harvests credentials from developer and continuous integration environments, including AI agent tokens, cloud keys, and self-hosted CI secrets, then uses stolen npm publishing access to poison more packages. A descendant of the earlier Shai-Hulud worm, it even forged valid-looking build provenance.
Manifold Security found 77 malicious extensions on Open VSX, the open marketplace that editors like VSCodium, Cursor, and Windsurf pull extensions from, impersonating legitimate developer tools. Detected between July 26 and August 1, all 77 were linked through a shared exfiltration domain and common code. Fifty-eight sent only basic system information, while the other 19 performed deeper reconnaissance, exfiltrating developer, Git repository, and continuous integration metadata. Manifold found no access to source code, credentials, tokens, or SSH material and did not determine the campaign's goal. The extensions were pulled from Open VSX by August 3, but must still be removed manually from affected machines.
Palo Alto Networks Unit 42 detailed a rebuilt version of XCSSET, macOS malware that spreads through Xcode developer projects. The actors compromise Git repositories and inject a downloader into ordinary files inside a project, so that when a developer builds the cloned project, a hidden loader runs under their own account with no permission prompt. It then infects every other Xcode project on the machine and propagates through shared code. Version 40 hides its logic in memory, generates polymorphic payloads, and adds a Chrome hijacker that drives the browser through its debugging protocol plus a component that trojanizes Telegram. Seventeen modules handle credential, keystroke, clipboard, and browser theft.
Palo Alto Networks Unit 42 described three techniques, collectively Pass-ta-key, in which malware already running on a compromised Windows PC abuses Google Password Manager's synced passkeys. None break the underlying cryptography; they target how Chrome stores device keys, re-enrolls a device, and whether a site verifies that a person was present. The techniques can silently obtain a valid login assertion without any unlock prompt, register an attacker-controlled verification key that defeats user-verification checks, or extract the secret that decrypts all of a user's synced passkeys for reuse elsewhere. Passkeys still resist phishing, but a fully compromised endpoint undermines the synced-passkey model.
Researchers at Socket found 18 malicious npm packages that deliver a cross-platform remote access trojan to users of Alibaba developer tools, splitting the attack across many packages so each looks harmless on its own. Ten lure packages with no real function depend on a bridge package, which pulls in loaders that fetch a rule-engine configuration from GitHub and use it to run OS-specific payloads from a server disguised as Alibaba infrastructure. On Windows it even replaces a legitimate Alibaba security app with a trojanized copy. The final trojan can steal data, run commands, and move laterally, and the campaign stayed hidden for about three months.
Resecurity reports that the INC ransomware operation has become the dominant group exploiting two SonicWall SMA1000 VPN appliance flaws, accelerating its attacks since early August. The pair, CVE-2026-15409 and CVE-2026-15410, were exploited as zero-days before SonicWall patched them in mid-July, and can be chained to gain root on the appliance and extract credentials, session databases, and one-time-password data. Many appliances remain unpatched or already compromised, leaving footholds attackers can reuse. Victims listed recently span private and government organizations across several countries, and some reported follow-up emails and phone calls from people claiming to help with the ransomware.