Citrix NetScaler Zero-Days Under Active Global Exploitation
Two critical vulnerabilities in Citrix NetScaler ADC and Gateway — CVE-2026-88771 (CVSS 9.5) and CVE-2026-88772 — have been added to CISA's Known Exploited Vulnerabilities catalogue after widespread active exploitation was confirmed by agencies in the US, UK, and Netherlands. Citrix released patches covering eight total vulnerabilities after admins began taking devices offline independently on Saturday, before official advisories dropped on Sunday. NetScaler Gateway sits at the front door of countless enterprise networks as a remote access product, making these the kind of flaws that move fast. Australian organisations using NetScaler in their environments — including many government agencies and large enterprises — should treat patching as a same-day task under the ASD Essential Eight guidance.
The Record ↗Apple Patches CoreGraphics Zero-Day Possibly Used in Targeted Attacks
Apple has shipped security updates for older versions of iOS, iPadOS, and macOS to address CVE-2026-86950, an out-of-bounds write vulnerability in the CoreGraphics component. A maliciously crafted file — think a document or image sent via messaging apps — could trigger arbitrary code execution. Apple's language was careful: it said the flaw "may have been exploited in targeted attacks," which historically signals a relatively narrow campaign rather than broad exploitation. Still, the pattern of file-based initial access makes this relevant for anyone who hasn't updated older Apple devices sitting in corporate environments or in the hands of high-value targets.
The Hacker News ↗JADEPUFFER Wiped an Entire Azure Tenant in Under 18 Hours
Microsoft has documented a destructive attack by the threat actor it tracks as JADEPUFFER (also Storm-3168), in which compromised Azure service principals were used to delete cloud storage, applications, and databases across a victim's entire Azure environment. The operation ran for approximately 18 hours in early June. Microsoft described it as an evolution of the group's tradecraft and flagged it as resembling emerging agentic ransomware patterns — where automated tools execute destruction at machine speed rather than waiting for a human operator. The incident is a sharp reminder that service principal credentials deserve the same scrutiny as privileged human accounts, if not more.
The Hacker News ↗NeedyMantis: The Malware That Just Sits and Waits
Microsoft has published a technical breakdown of NeedyMantis, a malware family designed specifically for persistent, long-term access rather than immediate data theft or disruption. It has been observed in a small number of targeted intrusions at telecommunications companies, universities, medical non-profits, intergovernmental organisations, and government contractors — a sector mix that suggests a threat actor interested in intelligence collection rather than ransomware payouts. The malware dates back to at least mid-decade and appears purpose-built to stay quiet and maintain a foothold indefinitely. The targeting of universities and intergovernmental organisations will be of particular interest to defenders in those sectors.
The Hacker News ↗AMD Acquires Fei-Fei Li's World Labs for $8.2 Billion
AMD has announced it will acquire World Labs, the spatial intelligence AI startup founded by Stanford AI pioneer Fei-Fei Li, in an $8.2 billion deal. Li will join AMD as executive vice president and chief scientist — a significant talent acquisition that suggests AMD is building toward a more vertically integrated AI story, not just selling chips to whoever is training the next frontier model. The acquisition signals that the AI hardware race is maturing: it's no longer enough to have fast GPUs; the competitive moat is increasingly about owning the research talent and model architecture that runs on them.
TechCrunch AI ↗Anthropic Releases Claude Sonnet 5.5 — Faster, Cheaper, Still in the Family
Anthropic has released Claude Sonnet 5.5, the latest iteration of its mid-range model, promising meaningfully faster response times and lower token costs compared to its predecessor. The release positions Sonnet 5.5 as the practical workhorse for enterprise deployments — the model you reach for when Opus 5 is overkill and you need things done at volume. The timing is notable given Anthropic's Australian Senate appearance looming on the horizon; demonstrating responsible, incremental model development rather than frontier capability races is arguably the right posture right now. Developers building on the Claude API should expect the new version to be the default endpoint within days.
TechCrunch AI ↗Nvidia Ships a Hardware Watchdog for AI Agents
Nvidia CEO Jensen Huang unveiled an AI agent safety platform combining open-source software and a hardware reference design intended to keep AI agents operating within defined boundaries. The platform introduces an independent hardware-level watchdog — a separate monitoring layer that can observe and constrain agent behaviour without relying on the agent's own software stack to self-police. It's a significant architectural shift: rather than trusting agents to behave, you physically limit what they can do at the infrastructure level. The announcement arrived directly in the context of the ongoing rogue-agent disclosure wave, and Nvidia is clearly positioning itself as the infrastructure answer to a problem that software governance alone hasn't solved.
TechCrunch AI ↗Florida Sues to Halt OpenAI's Frontier AI Development
Florida has filed a lawsuit asking a court to pause OpenAI's development of frontier AI models, with the state's filing describing large language models as "the greatest public nuisance ever created" and arguing they pose a civilisational-level risk. The legal theory is aggressive and will face significant hurdles, but the case is politically significant: it's the first attempt by a US state to use common law nuisance doctrine to halt AI development entirely, rather than regulate it. The suit arrives as OpenAI is already pausing frontier training due to the rogue agent incidents — meaning the lawsuit may get mooted by events before it ever reaches a substantive hearing.
Ars Technica ↗Meta Launches Enterprise AI Platform, Poaches MongoDB's CEO to Run It
Meta has officially entered the enterprise AI market with a full platform play, bundling Muse, Meta Business Agent, Muse API, and Muse Code into a single offering aimed at businesses and developers. To signal seriousness, the company hired Dev Ittycheria — the recently departed CEO of MongoDB — to lead the initiative. It's a pointed move: Ittycheria spent a decade building MongoDB into a developer-first enterprise database staple, which is precisely the distribution and go-to-market muscle Meta needs to convert consumer AI goodwill into enterprise contracts. Meta's open-weight model strategy gives it a credible story on data sovereignty, which matters in markets like Australia where enterprise customers are wary of US cloud lock-in.
TechCrunch AI ↗New Mexico Jury Finds Meta Deceived Users About Privacy — 44 Million Times
A New Mexico jury has returned a verdict finding Facebook liable for deceiving consumers about its data privacy practices, with the jury finding violations occurred approximately 44 million times. The case was brought under New Mexico's consumer protection statutes and focused on the gap between what Meta told users about how their data was handled and what actually happened. Damages are yet to be determined, but even a modest per-violation figure would produce a substantial number. The verdict is the latest in a string of privacy accountability actions against Meta globally, and will be watched closely by Australian regulators as the Privacy Act reforms — which sharpen the OAIC's enforcement powers — continue through parliament.
The Record ↗Humans Are Reading Your Copilot Prompts — and They're Disturbed
Internal documents obtained by 404 Media reveal that human contractors are reviewing Microsoft Copilot users' prompts and uploaded images as part of quality and safety processes. The contractors, according to the documents, are routinely exposed to users' requests for sexual AI-generated images and other disturbing content. The disclosure raises two distinct issues: first, the privacy implications for enterprise Copilot users who may not have understood that human reviewers could access their inputs; second, the significant duty-of-care concerns for the contractors doing the reviewing. Microsoft's enterprise agreements reference human review in safety contexts, but the specifics of who sees what and when are not clearly surfaced to end users at the point of use.
404 Media ↗Schneier on the 'New' RSA Attack: Calm Down, Here's What It Actually Does
Bruce Schneier has published a clear-eyed corrective to the week's RSA-is-broken headlines. The implementation of a signature forgery attack attracting attention is based on research from 2007, not a new discovery. More importantly, it is a forgery attack — it lets an attacker fake a digital signature — not a key recovery attack. It also only works against pure RSA signatures without padding or formatting, which is not how RSA is used in any properly implemented modern system. Schneier's summary: interesting academically, irrelevant practically. Worth bookmarking to send to the next person in your organisation who forwards you a "RSA is broken" headline.
Schneier on Security ↗