GitLab Code-Injection Flaw Under Active Exploitation Within Days
A critical code-injection vulnerability in GitLab — CVE-2026-19478, CVSS 9.4 — is already being exploited in the wild, just days after public disclosure. An unauthenticated attacker can modify or delete publicly accessible GitLab projects and rewrite their data without logging in. WatchTowr flagged the active exploitation, noting how quickly the window between disclosure and weaponisation has narrowed. Any organisation running self-hosted GitLab instances should treat this as an emergency patch, not a scheduled update. GitLab's prevalence in Australian developer infrastructure — including federal agencies and large enterprises running on-premises — makes prompt action critical.
The Hacker News ↗New Phishing Kit Hijacks Passkeys to Survive Password Resets
The promise of passkeys is that stolen passwords no longer matter. A new phishing toolkit called iAuthFlow V2 is stress-testing that promise. Researchers say attackers using the kit can register an attacker-controlled passkey on a victim's account during a phishing session — meaning that even after the victim notices the breach, resets their password, and revokes active sessions, the attacker's registered passkey remains a valid door in. It's a persistence mechanism that most users won't think to check for. Defenders should be reviewing whether their identity platforms allow users to audit and revoke registered passkeys, and whether anomalous passkey registration triggers any alerts.
SecurityWeek ↗North Korea Poisons Rust's arrayref Crate in Supply Chain Attack
North Korean hackers have added another programming ecosystem to their supply chain hit list. Researchers have attributed a poisoned update to the popular Rust crate arrayref — which has tens of thousands of downstream dependents — to the same threat cluster responsible for previous npm and PyPI attacks. The malicious update added a dependency that fetches a remote payload, effectively turning routine cargo build commands into a malware delivery mechanism. Rust's reputation for safety and rigour has made it a go-to for security-sensitive projects, which makes it an increasingly attractive target. Developers should audit their Cargo.lock files and verify the integrity of recent arrayref versions.
SecurityWeek ↗Cisco Crosswork and Secure Workload Hit With Five CVSS 10.0 Flaws
Cisco's latest security advisory reads, as The Register put it, like Olympic gymnastics scores: 10, 10, 9.9, 9.6, and 7.5. Five of the nine vulnerabilities patched in Crosswork Data Gateway, Crosswork Network Controller, Crosswork Planning, and Secure Workload Software carry the maximum CVSS score of 10.0. The flaws affect systems regardless of configuration, and SaaS-hosted customers also have updates to install — a detail that's easy to miss. Cisco's Crosswork platform is widely deployed in telco and large enterprise network operations. Australian telcos and managed service providers running these platforms should treat this advisory as a priority.
The Hacker News ↗Russian OAuth Abuse Targets Government Officials in Phishing Wave
Russian intelligence-linked actors are adding OAuth token abuse to their phishing playbook, targeting government officials and diplomats with convincing fake meeting invitations — including lures mimicking State Department communications. Rather than stealing passwords, the technique harvests OAuth tokens, granting persistent access to cloud accounts even when credentials are later changed. The Register notes this is a meaningful tactical evolution: OAuth abuse is harder to detect and revoke than credential theft. Organisations relying heavily on Microsoft 365 or Google Workspace for sensitive government communications should review OAuth application permissions and anomalous token grants.
The Register ↗CISA Flags Exploited TrueConf Bugs Deployed by Hacktivist Group
CISA has added TrueConf vulnerabilities to its Known Exploited Vulnerabilities catalogue after the Head Mare hacktivist group used them to deploy PhantomCore malware. TrueConf, a Russian-developed video conferencing platform sometimes called Russia's Zoom, has significant reach beyond Russia — it's deployed in enterprises and government agencies across multiple regions. The exploitation is being attributed to pro-Ukraine hacktivists, adding another layer of complexity to a product that some Western organisations had already quietly been auditing. Organisations still running TrueConf on-premises should patch immediately; those who haven't audited why it's on their network at all should probably start there.
SecurityWeek ↗OWASP Drops First Security Framework Specifically for AI Skills and Plugins
The Open Worldwide Application Security Project has published a dedicated top-10 risk list for AI skills — the modular plugins that let AI agents call external tools, APIs, and services. It also debuts a Universal Skill Format standard, aiming to bring consistency and baseline security requirements to how AI add-ons are built and distributed. The timing is pointed: agent-based AI is being deployed at pace in enterprise environments, often with limited scrutiny of what third-party skills can actually access. For Australian organisations building or procuring AI agent platforms, this framework is the closest thing currently available to a vendor-neutral security baseline.
Dark Reading ↗Nvidia Research: The AI Agent Harness Matters More Than the Model
Nvidia has published research suggesting that for agentic AI tasks, the scaffolding around a model — the harness — is now the dominant factor in performance and safety, not the raw capability of the underlying model itself. Fine-tuned, well-constrained agents using a modest model outperformed free-running frontier models on benchmark tasks, and critically, stayed within guardrails more reliably. The finding has practical implications for enterprise AI deployments: organisations investing heavily in prompt engineering, tool restrictions, and execution frameworks may be getting better safety outcomes than those relying on model capability alone to self-regulate. This also reframes the "agentic AI going rogue" problem — the harness is where the fight is won or lost.
TechCrunch AI ↗AI Models Designed Functional Bacteriophage Genomes — All 285 of Them
Bruce Schneier flags research that sits uncomfortably at the intersection of AI and biosecurity: two AI models were tasked with designing complete genomes for viable bacteriophages — viruses that infect and destroy bacteria — using an existing phage as a template. The models generated roughly 700,000 candidate designs; researchers selected 285 that appeared structurally viable. Schneier notes the dual-use tension plainly: the same capability that could accelerate therapeutic phage development for antibiotic-resistant infections could, with different intent, be pointed at more dangerous targets. Australia's Office of the Gene Technology Regulator and biosecurity frameworks are not currently calibrated for AI-generated genomic design at this scale.
Schneier on Security ↗Apple Cuts 200-Plus Jobs Across Vision Pro and Siri Teams
Apple has quietly laid off more than 200 employees spanning its Vision Pro and Siri organisations, according to Bloomberg. The Vision Pro gaming team has been largely shut down, and the immersive content team significantly reduced. Apple framed the cuts as "evolving the business" while promising new roles — the standard corporate softening. Read plainly, it signals a recalibration of the Vision Pro's ambitions: gaming and bespoke immersive content were apparently not the product's path forward. The Siri cuts are arguably more significant, coming as Apple continues to lag competitors in conversational AI. Apple products are ubiquitous in Australian enterprise and consumer markets, making its AI trajectory a question worth watching closely.
The Verge ↗Apollo Private Equity Discloses Breach in Wave of Financial Sector Attacks
Apollo Global Management — one of the world's largest private equity firms with hundreds of billions under management — has disclosed a data breach stemming from a social engineering attack on its cloud platforms. Attackers maintained access for a five-day window in early July, compromising sensitive personal data. Apollo joins a growing list of financial sector victims in what CyberScoop describes as an ongoing coordinated campaign against the industry. The breach is a reminder that high-value targets in finance face sophisticated social engineering, not just technical exploits. Australian financial institutions with US counterparty relationships or shared cloud infrastructure providers should assess whether they sit in the same exposure landscape.
CyberScoop ↗