Tuesday 15 September 2026

Perfect Score, Perfect Storm: The CVSS 10.0 GitLab Flaw Already Being Exploited in the Wild

A perfect-10 GitLab vulnerability is being actively exploited days after its patch dropped — and your CI/CD pipeline may be next.

Lead story

Perfect Score, Perfect Storm: The CVSS 10.0 GitLab Flaw Already Being Exploited in the Wild

A path traversal vulnerability in GitLab has earned the rarest rating in vulnerability scoring — a perfect 10.0 out of 10 — and attackers aren't wasting any time. CVE-2026-85706 affects both GitLab Community Edition and Enterprise Edition, and CISA has now confirmed it is being actively exploited. Watchdog firm watchTowr spotted threat actors probing internet-facing instances within days of the patch landing.

Path traversal bugs are among the oldest tricks in the book: by manipulating a file path request, an attacker can step outside the directory they're supposed to be in and access — or overwrite — files they have no business touching. In a platform like GitLab, which sits at the heart of software development pipelines, that means source code, secrets, environment variables, CI/CD configurations, and deployment keys are all potentially in play.

The supply chain implications are significant. GitLab isn't just a code repository — it's often the first domino in a chain that ends with production software being shipped to customers. A compromised GitLab instance is effectively a compromised software factory. We've seen that movie before with SolarWinds and 3CX: attackers who get inside the build pipeline don't just steal code, they insert things into it.

What makes this particularly sharp is the timeline. The patch has been available, but the gap between "patch lands" and "attackers exploit" has compressed dramatically over the past year. Security teams that treat patching as a weekly hygiene task are operating on attacker time, not defender time.

Who is at risk? Any organisation running a self-managed GitLab instance — CE or EE — that hasn't applied the patch. Cloud-hosted GitLab.com instances are managed by GitLab and were updated separately, but organisations running their own servers (common in enterprises, government, and regulated industries that keep code on-premises) need to act now.

Australian context: GitLab is widely deployed across Australian government and enterprise environments, where on-premises code hosting is common under data sovereignty and security policies. Agencies subject to the Australian Government Information Security Manual (ISM) and Essential Eight controls — particularly those under ASD's patch application requirements — should treat this as a Priority 1 remediation. The ACSC has not yet issued a specific advisory at time of writing, but given CISA's active-exploitation confirmation, one would be consistent with their standard response pattern.

What to watch: Red Heron — a suspected Chinese threat actor featured elsewhere in today's brief — has already been caught mass-scanning Gitea instances (a separate but related self-hosted Git platform) using a similar critical RCE flaw. The playbook of scanning first, exploiting at scale, and establishing persistent access is well established. It would be surprising if a CVSS 10.0 GitLab flaw didn't attract similar attention quickly.

The immediate action is straightforward: patch, then audit your GitLab logs for unexpected access patterns, particularly around file access outside standard project directories. If you're running GitLab in a CI/CD capacity, treat any unpatched window as a potential integrity gap and review recent pipeline outputs accordingly.

Also today

DDRop: The Hardware Attack That Makes Confidential Computing Less Confident

Researchers have disclosed DDRop, a hardware-level attack that silently drops memory writes to a server — leaving processors reading stale, encrypted data as if it were current. The attack undermines the memory protections built into Intel TDX and AMD SEV-SNP, the two dominant confidential computing technologies used to protect sensitive workloads in cloud environments. The catch: an attacker needs software control of the server and brief physical access to insert a small circuit. That limits real-world risk, but for high-value targets — financial data, government workloads, AI model weights — even a hardware supply-chain moment of access is plausible. Australian cloud-hosted sensitive workloads relying on confidential computing attestation should note this finding.

The Hacker News ↗

Red Heron Scanned 1,386 Gitea Servers Across Seven Countries in Multi-Nation Espionage Campaign

A suspected Chinese threat actor tracked as Red Heron has been attributed to mass exploitation of a critical remote code execution flaw in Gitea, the self-hosted Git platform. Acronis threat researchers found the group scanned nearly 1,400 internet-facing instances across seven countries and kept a separate, targeted list of 477 Taiwan-based systems — a geographic focus consistent with Chinese state intelligence priorities. Thirteen organisations across six countries were successfully compromised. For defenders, this is another data point on how quickly Chinese actors operationalise newly disclosed vulnerabilities at scale — and a reminder that Gitea, often running in smaller teams without enterprise security tooling, is very much in scope.

The Hacker News ↗

Revolut Handed Customer Data to Fraudsters Who Spoofed a Government Email

Revolut has confirmed it disclosed sensitive personal and financial data to fraudsters who submitted emergency data requests using a legitimate government email account — meaning the criminals didn't hack Revolut, they impersonated an authority that Revolut was legally obligated to trust. The social engineering attack exploited the emergency data request (EDR) process, a mechanism law enforcement agencies use to bypass standard legal channels in urgent situations. EDR abuse has become a significant vector for data theft. Revolut has Australian customers, and the incident raises questions about how fintechs operating under Australia's Privacy Act and OAIC notifiable data breach scheme verify the identity of government requestors before complying.

The Record ↗

JFrog Artifactory Flaws Exploited to Plant Backdoors in Software Build Systems

Three vulnerabilities in JFrog Artifactory — a widely used artefact repository that sits inside software build pipelines — are being actively exploited to bypass authentication and escalate privileges to administrator level. From there, attackers have been deploying backdoors. Artifactory is to build pipelines what GitLab is to source control: a critical junction where tampering can have downstream consequences for every software package that passes through it. The combination of GitLab and JFrog flaws both being exploited simultaneously is a bad week for software supply chain security. Australian DevOps teams using Artifactory in on-premises or self-managed configurations should treat patching as urgent.

SecurityWeek ↗

New York Seizes 12 Deepfake Websites in Largest-Ever Action of Its Kind

The Manhattan District Attorney's Office has seized 12 websites that collectively hosted sexually explicit deepfake imagery of approximately 1,200 people, including celebrities. The DA's office described the imagery as "hyper-realistic" and the action as the largest legal enforcement against harmful deepfake platforms ever undertaken in the United States. A separate Wired analysis found more than 100 female politicians across 22 European countries appearing on similar sites. Australia's Online Safety Act gives the eSafety Commissioner removal powers over non-consensual intimate imagery, including deepfakes — but seizure authority of the kind exercised in New York remains a distinct legal mechanism Australia does not yet have at equivalent scale.

404 Media ↗

Inside 'Project Lily': OpenAI Has Humans Reading Your ChatGPT Conversations

Leaked internal documents obtained by 404 Media reveal that OpenAI runs a programme called Project Lily in which human contractors read ChatGPT user prompts — including sensitive and personal ones — to improve model quality. The practice isn't unprecedented in AI development, and OpenAI's terms of service technically permit it, but the scale and the sensitivity of what users share with ChatGPT (medical questions, legal situations, relationship problems) makes the disclosure uncomfortable. It's a meaningful reminder that "your data helps improve our models" is not a hypothetical clause. Australian users should note this sits within OpenAI's US-based data processing, outside the reach of Australian Privacy Act protections.

404 Media ↗

Microsoft Publishes AI Code of Conduct — Telling Its Models Not to Hack or Deceive

Microsoft has released a formal "code of conduct" for its AI models, laying out principles including supporting humans rather than replacing them, and hard constraints like not hacking systems or deceiving users. The document is notable less for its content — most of it is motherhood-and-apple-pie safety language — and more for the fact that major AI vendors are now publishing formal behavioural specifications for their models at all. It's a governance trend worth tracking: as AI agents get more autonomy, the question of what they're explicitly prohibited from doing (and who enforces it) becomes increasingly consequential. Australia's AI Safety Standard consultation process is watching this space closely.

TechCrunch AI ↗

ConnectWise Patches ScreenConnect Flaw Being Exploited in Worm-Like Attacks

ConnectWise has patched a vulnerability in ScreenConnect — its widely deployed remote desktop and support tool — that allows attackers to send and execute files without authorisation through an active remote session. The flaw is being exploited in worm-like attack patterns, meaning compromised instances are being used to spread to further systems automatically. ScreenConnect is used heavily by managed service providers (MSPs) in Australia as a remote support tool, making any exploitable flaw in it a potential lateral-movement mechanism across multiple client environments simultaneously. MSPs operating under the ACSC's Essential Eight framework should treat this as a high-priority patch.

SecurityWeek ↗

UK Government Rolls Out Passkeys to 23 Million Citizens — and Saves £600 a Day on SMS

The UK's GOV.UK One Login identity system has begun migrating its 23 million users away from passwords toward passkeys, replacing SMS one-time codes with device-based cryptographic authentication. Beyond the security benefit, the UK government is saving approximately £600 a day on SMS costs alone — a neat argument for passkeys that procurement-minded bureaucrats can appreciate. The move is one of the largest government passkey deployments globally. Australia's myGov identity platform has been expanding its own digital identity capabilities under the Digital ID Act, and the UK's results will be closely watched as a comparable benchmark for large-scale government passkey rollout.

The Register ↗

NSA Reorganises Into Five Mission Centres, Including Dedicated Cyber and AI Divisions

The National Security Agency is undertaking a major structural reorganisation, creating five distinct mission centres — with dedicated units for cyber operations and artificial intelligence. The restructure is being executed quickly, which signals urgency from agency leadership. The creation of a standalone AI mission centre is particularly notable: it suggests the NSA views AI not as a tool that sits inside existing teams but as a domain requiring its own operational command structure. For allied signals intelligence agencies — including Australia's ASD, a core Five Eyes partner — NSA reorganisations typically have downstream influence on structure and doctrine.

The Record ↗

WordPress Adds Automated Security Scanning to Plugin Updates

WordPress has announced that every plugin update pushed through the WordPress.org update API will now be automatically scanned for security issues before distribution. Previously, new plugins were reviewed before joining the directory, but updates shipped continuously without the same scrutiny — a gap attackers have exploited by pushing malicious updates to legitimate plugins after initial approval. The change closes one of the more significant open doors in the WordPress ecosystem. Given that WordPress powers roughly 43% of the web globally — including a substantial portion of Australian small business and media sites — this is a meaningful systemic improvement, even if automated scanning isn't a perfect substitute for manual review.

The Hacker News ↗

Sources consulted