ACSC Flags Global CMS Exploitation Wave as AI Research Questions Its Own Direction

Two distinct currents ran through yesterday's news: an escalating, automated campaign against CMS infrastructure that demands immediate action, and a quieter but increasingly vocal reckoning within AI about what the field is actually doing and who is steering it.

Security

The Australian Cyber Security Centre published an advisory warning of a global exploitation campaign targeting vulnerable CMS platforms and plugins. The campaign is opportunistic and automated — attackers are scanning for unpatched WordPress, Joomla, and other CMS installations at scale, chaining plugin vulnerabilities to achieve remote code execution, credential harvesting, and in some cases persistent backdoor access. The ACSC's framing makes clear this isn't a nation-state targeted operation but a broad criminal effort monetizing any reachable surface. That distinction matters: with targeted attacks you might reason about whether you're interesting enough to attract attention; with automated campaigns, the only question is whether your software is outdated. Every unpatched instance gets harvested regardless of size or content.

This advisory lands in the same window that already includes a Linux kernel UAF (GhostLock), a stored XSS in Zimbra, and the jscrambler supply chain compromise. The pattern isn't coincidence — it reflects a threat environment where multiple simultaneous campaigns are running across different layers of the stack. The CMS campaign is the widest in surface area, since plugin ecosystems are notoriously slow to patch and collectively power a large fraction of the public internet. If you're maintaining CMS infrastructure, audit your installed plugins against the ACSC indicators now and disable anything unused. Patch windows in automated campaigns are measured in hours.

AI

The Information's piece on AI researchers having an identity crisis is behind a paywall, but its existence and framing reflect something real: the people building these systems increasingly can't agree on whether they're doing science, engineering, or product work, and that confusion is generating friction at major labs. The era of capabilities-above-all benchmarking is colliding with the reality that the remaining hard problems don't fall neatly into any discipline.

George Hotz's AI 2040 takes a sharper line: intelligence is being commoditized, moats are collapsing, and the cult built around AI is the product now, not the underlying capability. He's been consistent on this for years, and reading it alongside the researcher identity crisis piece makes the dynamic concrete — the field is in a transition nobody fully controls. The governance question surfaced separately: Off-Policy's essay on agentic AI asks who actually manages the agents as they proliferate, not philosophically but practically. When agents are taking real-world actions across systems, the failure modes are no longer confined to bad outputs; they're systemic. The accountability question keeps getting deferred at every layer.

On the product front, OpenAI is hiring a dedicated PM to build ChatGPT experiences for families, caregivers, and older adults. The household play signals a bet that enterprise growth is plateauing and the next frontier is embedding into domestic life — a sharp contrast with the existential uncertainty happening at the research level. Separately, OpenAI forked Git on GitHub with minimal explanation. A company-maintained Git fork implies internal workflow needs that diverged far enough from upstream to make the fork cheaper than upstreaming patches — worth watching the commit log for what surfaces.

On the infrastructure side, Mesh LLM proposes distributing large model inference across peer-to-peer nodes using the iroh library — treating inference as something spreadable across heterogeneous consumer hardware rather than concentrated in datacenter GPUs. And a detailed writeup on running Qwen3.5-122B as a daily driver on Mac Studio covers three bug fixes that made a 122B parameter model usable at real speed on Apple Silicon. The practical upshot: frontier-scale models are increasingly runnable locally with the right tooling, quietly undermining the argument that serious inference requires datacenter access.

Tech

Firefox hit 12.58% desktop browser market share in North America for June 2026, per StatCounter. That number has been grinding downward for years, so a reading above 12% is worth flagging — though one month's data doesn't establish a trend. Firefox remains the only major browser independent of both Google and Apple, which gives the number significance beyond market dynamics alone.

The Nvidia/CoreWeave/Nebius circular financing story deserves a careful read: the piece traces how GPU revenue is recycled through cloud credit agreements, equity stakes, and lending relationships among these three entities in ways that inflate valuations on all sides by creating the appearance of demand that is, in part, self-reinforcing. It's the kind of structural analysis that's easy to dismiss in a bull market and very obvious in retrospect after corrections.

On the engineering side, ClickHouse published how they scaled PgBouncer to 4x throughput building their managed Postgres offering — practical reading for anyone running connection pooling at scale. And SLA Credit Watch launched as a public ledger of cloud outages and the credits they actually trigger, which is the kind of provider accountability infrastructure that's been conspicuously absent.

Martha Lillard, the last American polio patient living inside an iron lung, died at 78 in Oklahoma. She spent decades in the machine that kept her alive since childhood. Her death closes a chapter and opens a reminder.

The ACSC advisory is today's action item — patch before you read anything else.

Also yesterday

• GhostLock: 15-Year Kernel UAF Affects Every Linux Distribution
• Critical Zimbra Stored XSS Lets Crafted Emails Execute Code in User Sessions
• Compromised jscrambler 8.14.0 Drops Rust Infostealer on Install — Act Now