📰 The CyberSignal Daily — August 26, 2026

Tuesday was a day about the distance between a fix and its installation. CISA forced federal agencies onto the tightest patch clock it has ever set for a perfect-10 Oracle WebLogic flaw that Oracle fixed back in January, Shadowserver counted 274 compromised Zimbra servers running a version patched in July, and ANY.RUN published the receipts on a phishing-as-a-service kit that quietly worked roughly 4,500 US and EU companies for two years by letting Microsoft 365 logins succeed before stealing the session. Meanwhile Oasis Security showed the newer problem: an NVIDIA NemoClaw weakness where there is no patched version to install at all, because the exposed component is the unauthenticated model API sitting on a developer's own machine. And Treasury opened an economic front, sanctioning Iran-linked hackers tied to critical infrastructure intrusions. Patches, when they exist, are not the bottleneck. Getting them on the box is.

Vulnerabilities & Patching

CISA added CVE-2026-21962 to the Known Exploited Vulnerabilities catalog on Monday, citing active exploitation. The flaw is an improper access control weakness in Oracle HTTP Server and the Oracle WebLogic Server Proxy Plug-in, rated CVSS 10.0, exploitable by an unauthenticated attacker over HTTP with no user interaction. Federal civilian agencies must remediate by August 27 under BOD 26-04, a three-day window and the shortest the agency has ever issued.

The number that matters is not the 10.0. Oracle shipped the fix in its January 2026 Critical Patch Update, which means roughly seven months passed between a patch existing and CISA having to compel its installation, while unauthenticated exploitation ran against exposed servers. If you run WebLogic or Oracle HTTP Server, apply the January 2026 CPU now and treat any unpatched internet-facing instance as suspect rather than merely vulnerable. Hunt backward, not just forward.

→ Read the full story

Mail Server Security

The Shadowserver Foundation reported Monday that at least 274 internet-facing Zimbra Collaboration Suite instances have been compromised through CVE-2026-73570, a code injection flaw Synacor patched in ZCS v10.1.20 on July 20, 2026. The attackers remain unattributed. Zimbra is heavily used by organizations that need data control or cannot absorb Microsoft 365 and Google Workspace pricing, which concentrates the exposure in exactly the places with the thinnest patching capacity.

Mail servers are a compromise you cannot detect by patching alone, because a mailbox that was read stays read. Move to v10.1.20 or later immediately, then run an actual compromise audit: look for unexpected webmail extensions, unfamiliar admin accounts, modified server-side filters that forward mail externally, and new cron or startup entries under the Zimbra service account. Rotate service and admin credentials afterward.

→ Read the full story

Phishing & Identity

ANY.RUN published research on Mirage2FA, a commercial phishing-as-a-service toolkit that ran against roughly 4,500 US and EU companies between 2024 and 2026, with US-based organizations making up the bulk of the targets. The kit abuses legitimate Microsoft 365 login flows: the victim completes a genuine sign-in, including the second factor, and the operator captures the resulting authenticated session. ANY.RUN assesses that 48% of targeted email addresses were potentially compromised.

This is the failure mode that makes "we have MFA" a misleading control statement. Adversary-in-the-middle kits do not break the second factor, they wait for it and take what it produces. The defense is session-level, not prompt-level: enforce phishing-resistant methods such as passkeys or FIDO2 for privileged accounts, apply conditional access policies that bind tokens to compliant devices, and shorten refresh token lifetimes so a stolen session expires before it is useful.

→ Read the full story

Artificial Intelligence

Oasis Security disclosed a weakness in NVIDIA NemoClaw that lets an attacker-controlled webpage reach the local Ollama instance serving an AI agent and write hidden instructions into the model itself. The mechanics are unglamorous: the daemon binds to every network interface, and the model API on port 11434 requires no authentication, so a page loaded in the user's browser can reach it. Oasis reported the findings to NVIDIA's PSIRT. No confirmed CVE identifier or patched version exists at publication, so operators cannot check whether their installation is in scope.

Persistence inside a model is a category defenders do not have tooling for. Nothing on disk looks wrong and no process looks anomalous, but the agent's behavior is now attacker-influenced on every subsequent run. Until a fix lands, restrict the daemon to loopback, enforce Host and Origin allowlisting, lock down CORS, and block any non-loopback binding on machines running local agent stacks.

→ Read the full story

Nation-State & Policy

The US Department of the Treasury announced sanctions on Iranian cyber actors, including individuals connected to the Mabna Institute hacking-for-hire operation, describing the action as part of an "unprecedented, whole-of-government, economic campaign" against Iran and its enablers. Officials framed the objective as severing the financial connections that sustain the regime rather than deterring any single intrusion set.

Sanctions do not stop an intrusion in progress, but they do change the compliance surface for defenders. Designations mean new blocked-party screening obligations, and for critical infrastructure operators they are also a targeting signal worth acting on. Refresh your indicator coverage for Mabna-associated tradecraft, which historically leaned on credential phishing against research and academic accounts, and verify that vendor and payment screening lists have picked up the new designations.

→ Read the full story

Three of yesterday's four vulnerability stories share a shape that should be uncomfortable for anyone who measures security by patch availability. Oracle fixed CVE-2026-21962 in January. Synacor fixed CVE-2026-73570 in July. In both cases the fix existed, was public, and was ignored long enough for exploitation to become routine, and in both cases the correction arrived as an external forcing function: a federal mandate in one, a third-party scan count in the other. Vendors are not the constraint anymore. The constraint is the operational distance between a released patch and a restarted service.

The Mirage2FA research sharpens the same point from the identity side. Nothing in that campaign required an unpatched anything. It required only that organizations treat multi-factor authentication as a binary they had already checked off, while the attacker collected the artifact MFA produces rather than the credential it protects. Two years of quiet operation against 4,500 companies is not a technology gap. It is a measurement gap, where the control being reported ("MFA enabled") is not the control being tested ("session cannot be replayed elsewhere").

Then NemoClaw, which is the version of this problem with no patch to be late on. There is no confirmed CVE and no fixed version, so an operator cannot even determine exposure, and the persistence mechanism lives inside model weights where no endpoint agent is looking. Local AI stacks are shipping with defaults from a world where "localhost" meant "trusted," and browsers have been able to reach localhost for a decade.

The through-line for the week: your patch SLA is a claim about your organization, not about your vendors, and your MFA posture is a claim about sessions, not logins. Both are testable this morning. The AI agent running on your developers' laptops is probably not covered by either test yet.

Stay sharp. Stay ahead.

Till next time,

The CyberSignal Team

Our News Site | Get the Weekly Briefing