Researchers at ANY.RUN documented a phishing campaign spanning 46 countries, with about 45 percent of activity aimed at the United States, that tricks victims into installing legitimate remote monitoring and management software to give attackers persistent access. The lures pose as tax documents, invoices, shipping notices, and government messages, and the operation leans on disposable infrastructure hosted on trusted platforms like Vercel, GitHub Pages, and Netlify, with most hosts appearing for only a single day. Because the specific domains and remote-access tools are interchangeable while the delivery chain stays stable, defenders cannot rely on individual indicators or malware verdicts alone. Remote-management governance is the more durable control.
Group-IB detailed BraZetsu, a modular malware framework that turns compromised Windows machines into products sold to other criminals. It uses generative AI to triage stolen data and flag high-value victims for initial-access brokers, and it collects digital certificates, browser histories from several browsers, and financial files while watching users through screenshots. Compromised hosts feed an underground access-as-a-service marketplace where buyers can pay a small deposit to purchase entry into a victim's system and then run their own follow-on payloads. Some samples were fully undetected by antivirus at the time of analysis. It shows attackers using AI to scale the triage and resale of stolen access.
Researchers at Manifold Security disclosed a class of flaws across several command-line AI coding agents in which a repository's own Git configuration names a command that the agent runs automatically on the developer's machine. The command executes outside the agent's sandbox, with the user's privileges, and without any approval prompt, often before the agent even contacts the model. Simply reviewing or opening a malicious project can run attacker code. It triggers when a repository arrives as files with its hidden Git directory intact, such as through a shared drive, archive, or USB stick, rather than a normal clone. Several tools shipped fixes, but some remained vulnerable at disclosure.
Attackers used a BGP hijack, a manipulation of internet routing, to divert update traffic for Virtualizor, a widely used server and hypervisor management panel, to a server they controlled. During the diversion, which began August 28, they obtained a valid TLS certificate so the connection looked legitimate, then delivered a malicious update that installed persistent root access on affected hosts. One hosting provider found root-level compromise on five of thirty-four hypervisors it checked. Because the software's updates were not cryptographically signed, transport encryption alone did not stop the tampering once routing was hijacked. The vendor released a scanner and patch, but package signing remains unfinished, leaving update integrity dependent on routing security.
Researchers at GuidePoint found a ClickFix campaign that compromised at least 31 organizations' websites and abuses the Polygon blockchain to run its command-and-control, a technique called EtherHiding. Visitors arriving from search engines hit a fake human-verification prompt that abuses Cloudflare's overlay and tells them to paste a command, which installs a persistent backdoor. Instead of a fixed server address that defenders can block, the backdoor fetches its current instructions from a Polygon smart contract every minute, giving the attacker a censorship-resistant, easily updated address book. This breaks the usual defense of blocking a hardcoded command server, so defenders should focus on behavior and audit their public-facing sites.
A researcher known as Chaotic Eclipse, or Nightmare Eclipse, publicly released two unpatched privilege-escalation zero-day exploits targeting security software, without coordinating with the vendors. One, called HardBreacher, targets Kaspersky Endpoint Security and can disrupt the antivirus and its file-access controls while creating a system-level file. The other, PrettyPrague, escapes the Avast sandbox to dump the Windows account database and spawn a SYSTEM-level shell, and reportedly works on fully patched Avast and Windows 11. The researcher suspects it may also affect other Gen Digital products like AVG and Norton. Because there are no CVEs or patches yet, endpoints are exposed, and security tools' high privileges make them valuable targets.
ESET found that a Russia-aligned group planted a prompt inside a malicious script designed to trip an AI system's safety filters and disrupt AI-assisted malware analysis. The technique, dubbed GuardBreaker, embeds text about a sensitive topic so that a language model reviewing the code refuses or derails instead of analyzing it. The script itself installs a loader the group uses to deliver further payloads. It is not isolated: earlier in 2026, malicious packages in supply-chain campaigns used similar anti-analysis tricks against systems leaning on a language model for triage. The lesson is that automated AI triage can be manipulated by the code it inspects, so human review remains essential.
Researchers found thirteen malicious packages on Packagist, the PHP Composer registry, posing as content-management themes that inject JavaScript into the sites that use them. On visitors' devices the script runs gambling and ad-fraud redirects, and on iPhones it loads a WebKit exploit chain that, against unpatched devices, installs spyware and steals cryptocurrency wallet seed phrases. The packages span several vendor names and extend a campaign first seen in March that abused similar theme packages and attacker-hosted infrastructure. It is a reminder that a compromised server-side dependency can become a delivery system for attacks against every visitor, including mobile users, not just the server it runs on.
Researchers at CloudSEK and Gambit Security found that operators of the Russian-speaking Aurora ransomware used the agentic coding assistant Cursor to help break into around ten victim networks. After obtaining valid credentials or a route in, the operator directed the agent to run reconnaissance, assess privileges, scan internally, and attempt exploitation, often revising commands several times before they worked, which shows the human stayed in control and used the AI as an assistant rather than an autonomous attacker. The group also built a Linux encryptor that force-terminates VMware ESXi guest virtual machines to unlock their disk files before encrypting them. Initial access in one case came through help desk impersonation phone calls.
The AI evaluation nonprofit METR disclosed that attackers stole a model-provider API key and ran up about 600,000 dollars worth of inference credits over three weeks. The key sat on a researcher's personal cloud instance that was meant to be protected by a Google login but, due to a fail-open authentication bug in a quickly built app, was actually publicly reachable. After finding it, the attacker prompted the AI agent running there to reveal its provider API key, added an SSH key for persistence, and consumed credits on public models. The abuse went unnoticed for a while because METR routinely runs high-token evaluations and had no spending caps on the key.