☀️ Good morning. Here's everything that happened in cybersecurity yesterday, in under 5 minutes.
Attackers began hitting a critical GitLab flaw roughly two days after the fix shipped, while the second WordPress plugin in three days turned an ordinary contact form into unauthenticated remote code execution. An AI coding agent handed a developer the name of a malicious package, and a company rule about reading the source on GitHub first was the only thing that stopped the install. Researchers at UMass Amherst showed a contactless card can keep making payments past its printed expiration date, even after the replacement arrives. Our weekly wrap supplies the numbers behind all of it: Rapid7 logged 8,539 high- and critical-severity disclosures in a single quarter, and the share of new flaws seen exploited jumped from 8% to 40%.
🔥 Top Stories
01 — Attackers Exploit Critical GitLab Flaw Two Days After the Patch
Vulnerabilities
WatchTowr's honeypot network logged the first in-the-wild exploitation attempts against CVE-2026-19478, the unauthenticated GraphQL vulnerability GitLab patched on August 17, SecurityWeek reported on August 20. The flaw carries a CVSS score of 9.4 and lets an attacker with no credentials remotely modify or delete public projects and user data on self-managed Community and Enterprise instances. GitLab shipped fixes in 19.2.4, 19.1.6, 19.0.8, and 18.11.11. WatchTowr researcher Jake Knott said a single HTTP request is enough to delete repositories, forge merge records, or ban maintainers, with no user interaction or unusual configuration required.
The timeline is the story: patch on August 17, a working exploit rebuilt by WatchTowr on August 18 from the advisory and patch diff alone, real attempts landing in honeypots by midweek. Patch to a fixed build on your release branch now. If you cannot, restrict unauthenticated access to the /api/graphql endpoint or remove public repository access until the update is staged, then search web and application logs for requests containing @gl_introduced and review merge and approval records for forged entries. The CVE is not on CISA's KEV catalog as of publication, so do not wait for a listing to act.
02 — Elementor Pro Flaw Lets Unauthenticated Visitors Upload PHP for RCE
Vulnerabilities and Patching
Patchstack disclosed CVE-2026-32475, a CVSS 9.0 unrestricted file upload flaw in the Elementor Pro Forms module that lets an unauthenticated visitor drop a PHP file on the server and execute it. The extension check and the file-move step run in two separate loops that disagree about how to handle empty file entries, so the blocklist meant to reject PHP gets sidestepped and the file lands in wp-content/uploads/elementor/forms/. Every version up to and including 4.2.1 is affected; the fix shipped in 4.2.2 on August 19. Only the paid Pro add-on is affected, not the free Elementor page builder.
The precondition is trivial: one published page with a Form widget that includes a File Upload field, a setup Patchstack calls "an extremely common, everyday configuration." This is the second unauthenticated upload-to-RCE flaw in a mainstream WordPress plugin in about three days. Update Elementor Pro to 4.2.2, then audit wp-content/uploads/elementor/forms/ for any .php files, since nothing executable belongs in that path and a hit there is a strong indicator of compromise. If you cannot patch tonight, delete the File Upload fields from published forms. No public exploitation has been reported yet, and the bug is not on KEV.
03 — An AI Coding Agent Recommended Malware. A GitHub Habit Stopped It.
Artificial Intelligence and Supply Chain
An engineer at software consultancy Softjourn asked an AI coding agent to recommend a package for a routine task, and the agent returned a legitimate-sounding name that turned out to be malicious, The Register reported on August 20 in its PWNED column. Company policy required a human to review any AI-recommended package on GitHub before installing. The engineer did, saw a package with almost no downloads that had been created only days earlier, and stopped the install. The pattern has a name: slop-squatting, where attackers register the exact package names large language models tend to hallucinate and fill them with malicious code.
Slop-squatting is typosquatting's harder cousin. Typosquatting bets on a human fat-fingering a real library name; slop-squatting bets on machine confidence, and the model volunteers the target list. The risk sharpens as agents gain permission to run install commands on their own, removing the pause where a person would notice. Make the pause mandatory: require a source review covering package age, download history, maintainers, and recent commits before anything enters the project, keep a human approval gate between an agent's suggestion and a dependency landing in the build, pin and allowlist dependencies, and turn on registry cooldown windows that hold back brand-new versions. The Register did not name the AI tool, the package, or the registry.
04 — Expired Contactless Cards Can Still Make Payments, UMass Amherst Shows
Payments Security
A team at the University of Massachusetts Amherst demonstrated that a contactless credit card can keep making payments after the expiration date printed on its face, and in their tests it kept working even after the bank issued a replacement. Presented at USENIX Security 2026 as "Zombie Cards Back Online," the work used a relay built from two ordinary smartphones to present a future expiration date to the point-of-sale terminal. In the affected Visa configuration that date is not covered by the card's cryptographic checks. Against one major US bank the reworked card cleared test purchases of $1, $100, and $500, then live purchases of $2.79 and $3.19 at retail and grocery checkouts. Mastercard, American Express, and Discover configurations rejected the altered date, and a second issuing bank declined the tap.
The weakness is an ownership gap, not a broken algorithm. A contactless payment passes through the card, terminal, merchant bank, network, and issuing bank, and each holds only a piece of the expiry decision; when verification confirms only that the account exists and the number is active, a replaced card is not treated as any more dead than an expired one. Visa and the affected banks were notified in 2025, but no fix was confirmed by publication and no CVE has been assigned. For cardholders, the cheap defense is the oldest one: physically destroy expired cards through both the chip and the number, and watch statements on closed and replaced accounts. For issuers and networks, verify the specific card and expiry on file rather than the account alone.
05 — Weekly Wrap: Iran Charges, a Citrix Pre-Auth Bypass, and Exploits Outrunning Patches
Weekly Roundup
The Justice Department unsealed a superseding indictment charging 17 members of Iran's Mabna Institute over a campaign that stole roughly 31.5 terabytes of academic data and intellectual property across 144 US universities, 178 foreign universities, dozens of private companies, and several US government agencies, largely through password spraying and credential theft. Citrix patched CVE-2026-19490, a CVSS 9.3 pre-authentication bypass in NetScaler ADC and Gateway that SecurityWeek reports is expected to be exploited; fixed builds are 14.1-73.32 and later and 13.1-63.21 and later. Firefox 154 patched 58 CVEs including 20 rated high, and Chrome 151 fixed 15 issues including a critical buffer overflow.
Rapid7's Q2 numbers put a figure on the squeeze: 8,539 high- and critical-severity disclosures, roughly double the 4,268 from a year earlier, with the exploited share of newly disclosed flaws rising from 8% to 40%, public proof-of-concept code up 76% year over year, and about 62% of newly exploited flaws requiring no authentication or user interaction. Elsewhere: Grandoreiro resurfaced with Mexico accounting for about 40% of detections via DLL sideloading, eSentire detailed a malware-as-a-service chain that stacks ClickFix lures with a loader that abuses a signed driver to kill EDR from the kernel, and Check Point mapped an operation running close to 2,000 hacked WordPress sites as attack infrastructure. Prioritize by real exposure and known exploitation rather than CVSS alone, enforce MFA on every external login, and deploy Microsoft's vulnerable-driver blocklist.
📊 By The Numbers
CVSS 9.4 — Severity of GitLab's CVE-2026-19478, an unauthenticated GraphQL flaw that can delete or rewrite public projects in a single request.
~72 hours — Elapsed time from GitLab's August 17 patch to real exploitation attempts landing in WatchTowr's honeypots.
4.2.2 — The Elementor Pro build that closes a CVSS 9.0 unauthenticated file-upload path to remote code execution. Every version through 4.2.1 is exposed.
8,539 — High- and critical-severity vulnerability disclosures Rapid7 logged in Q2 2026, roughly double the 4,268 from a year earlier.
31.5 terabytes — Academic data and intellectual property prosecutors say Iran's Mabna Institute stole, in a case that now charges 17 people.
⚡ The Signal
Four of yesterday's five stories describe a control that existed and did not get applied in time. GitLab shipped a patch and attackers arrived two days later. Elementor's blocklist was there, but two loops disagreed about empty entries and a PHP file walked through the gap. Visa's cryptographic checks cover most of a contactless transaction, just not the expiration date the whole system treats as authoritative. The AI package recommendation was caught by a policy, and only because someone actually followed it. In each case the security engineering was not missing. The enforcement was.
Rapid7's quarter is what that looks like at scale. When the exploited share of newly disclosed flaws climbs from 8% to 40% in three months and 62% of those require no authentication, the assumption underneath most patch calendars, that disclosure buys you a comfortable window, stops being true. A severity-ranked queue processed on a maintenance schedule is a plan built for a slower adversary. What replaces it is unglamorous: an inventory of every internet-facing instance and its version, an emergency lane that ignores the monthly cycle, and prioritization driven by exposure and known exploitation rather than a CVSS number alone.
The AI thread runs the same direction, and it is worth being precise about why. Slop-squatting did not invent a new class of vulnerability. It removed a pause. The developer who copies a package name from a chat window still has a moment to check; the agent wired to install on its own does not, and it will act on a hallucinated name with the same confidence it brings to a correct one. Softjourn's save came from a person spending two minutes on a GitHub page, not from a scanner. As attackers industrialize the easy parts, defenders have to industrialize the boring ones first, which mostly means making the checks that already work impossible to skip.
🔍 What You May Have Missed
The CyberSignal published five articles yesterday, all featured above, so these three come from the wire.
Rust Supply Chain Attack Puts Build-Time Malware in Crates With 245 Million Downloads — The Rust Project deleted malicious versions of arrayref 0.3.10, internment 0.8.7, and append-only-vec 0.1.9 from crates.io after a compromised maintainer account added a typosquatted dependency whose build script fetched and ran a remote payload at compile time. Check your lockfiles for those exact versions.
Suspected Russian Hackers Abuse Google OAuth and WhatsApp Linking to Hijack Accounts — Google Threat Intelligence is tracking three clusters (UNC6293, UNC7005, UNC5976) abusing legitimate authentication flows against academia, aerospace and defense, governments, and think tanks in Europe and the US. Device-linking and OAuth consent abuse bypasses the phishing page entirely.
Attackers Exploit Zimbra SNMP Flaw for Unauthenticated Remote Code Execution — CERT Polska reports active exploitation of CVE-2026-73570, a CVSS 8.9 command injection in Zimbra Collaboration. The flaw is patched, so this is a patch-status check, not a research problem.
📅 What to Watch
New Cryptographic Context Injection Attack Could Let Web Pages Steal Grok Chat Data — Adversa AI says an ordinary "summarize this page" request can push a user's name, approximate location, subscription tier, and current conversation to an attacker-controlled server. Watch for xAI's response and for the technique being retried against other assistants.
NASA AIT-GUI Flaws Could Let Unauthenticated Attackers Issue Spacecraft Commands — Cycode disclosed a CVSS 9.4 chain (GHSA-p9r8-2q67-fp86) in the operator console for NASA/JPL's open-source AMMOS Instrument Toolkit. Any organization running AIT-GUI should be tracking the fix.
US Bank Investigates LockBit's Claims as Ransomware Crew Sets Pay-or-Leak Deadline — The bank says it is investigating. The deadline gives this one a clock, and whether data actually posts will say a lot about the claim's credibility.
The Push to Designate AI as the Next Critical Infrastructure Sector — A designation would unlock federal services and resources for the sector, and would pull AI providers into the same reporting and resilience expectations as power and water. Expect the policy fight to sharpen after this week's Siemens S7 advisory.
Stay sharp. Stay ahead.
Till next time,
The CyberSignal Team