Daily brief at 7am Melbourne. Unsubscribe any time.

Wednesday 15 July 2026

622 Patches in One Day: Microsoft's AI Vulnerability Machine Is Running Hot

Microsoft's AI-powered vulnerability scanner just broke its own record — twice in two months — and two of the 622 flaws are already being exploited in the wild.

Lead story

622 Patches in One Day: Microsoft's AI Vulnerability Machine Is Running Hot

Last month, Microsoft's June Patch Tuesday dropped around 206 CVEs and everyone called it unprecedented. Yesterday, the company more than tripled that figure — shipping fixes for 622 vulnerabilities in a single release. Two of them were already being actively exploited before the patches landed.

The two zero-days are the ones defenders should move on first. One sits in Active Directory, the other in SharePoint Server. Microsoft says incident responders flagged both, which means attackers had already found them in the wild and were using them before any fix existed. A third flaw — in BitLocker — was publicly disclosed before yesterday's release, giving potential attackers a head start there too.

The sheer volume of this release is the real story. Microsoft has been open about the reason: the company is now using AI to systematically scan its own codebase for vulnerabilities, and the machine is finding them faster than any human team could. The June record of ~206 was itself a spike; 622 in a single month represents an exponential jump. CyberScoop quoted Microsoft warning customers that a flood was coming. It delivered.

This raises a genuinely uncomfortable question. If AI is surfacing vulnerabilities at this rate, the backlog inside large software estates — not just Microsoft's — may be far larger than the industry previously assumed. Every patch that ships is a good thing. But the remediation burden on enterprise security teams is growing faster than most patching programmes can absorb.

The practical advice for defenders is straightforward, even if the workload is not: prioritise the two actively exploited zero-days (Active Directory and SharePoint), then the publicly disclosed BitLocker flaw, then tier the rest by exposure. The CVSS scores and Microsoft's own exploitability ratings are your triage guide. Anything internet-facing or handling authentication comes first.

For Australian organisations, this is a significant patch cycle. Active Directory and SharePoint are deeply embedded in Commonwealth and state government environments, as well as enterprise Australia broadly. The Australian Signals Directorate's Essential Eight maturity framework includes patching as a core control — but at 622 CVEs, even organisations sitting at Maturity Level 3 will need to think carefully about prioritisation workflows. The ASD's ACSC advisory feed is worth watching for any follow-on guidance specific to Australian critical infrastructure sectors.

The bigger pattern is worth stepping back to consider. We are entering an era where AI doesn't just help attackers move faster — it also helps defenders (and vendors) find flaws faster. The question is whether the defensive side of that equation scales quickly enough. Right now, the answer is: not obviously. Security teams are already stretched. A tripling of patch volume in a single month is a stress test most weren't ready for.

Watch for whether other major vendors — Google, Oracle, Adobe — follow with similarly inflated patch counts as they adopt comparable AI scanning. If this is an industry-wide shift, yesterday's "patchpocalypse" may become the new normal.

Also today

Australia Is Getting a Government AI Office

The Australian federal government has announced it will establish a dedicated AI office to coordinate regulation across departments. The move signals Canberra is moving past ad-hoc AI governance and toward a centralised oversight structure — something the tech sector and civil society groups have been pushing for. Details on the office's mandate, staffing, and relationship to the existing National AI Centre remain thin, but the direction is clear: Australia wants a single point of coordination as AI policy becomes too complex for any one agency to handle alone. The announcement puts Australia roughly in step with the EU's AI Office and the UK's AI Safety Institute, though the scope and powers will matter enormously.

iTnews

New York Becomes First US State to Freeze Data Centre Construction

New York Governor Kathy Hochul has signed a one-year moratorium on new large-scale data centre approvals — the first such ban by any US state. The move is a direct response to the AI infrastructure boom, with Hochul arguing that unchecked data centre growth is driving up electricity costs for ordinary residents, straining water supplies, and bypassing local planning controls. The moratorium rattled the AI industry, which had earmarked New York as a key expansion corridor. If it survives legal challenge, it could become a template for other states looking to slow the build-out. Australia faces analogous tensions, with data centre developments in NSW and Victoria already drawing scrutiny over energy and water use.

Ars Technica

DeepMind CEO Wants a FINRA for AI

Demis Hassabis has called for an independent international standards body to test and certify frontier AI models before they are released — modelled explicitly on FINRA, the US financial industry's self-regulatory organisation. The proposal, made publicly this week, would have the body develop evaluation benchmarks, audit model behaviour, and set release conditions. It is a notable departure from the voluntary commitments that have defined AI governance to date. Hassabis argued that the pace of capability development has outrun any individual government's ability to assess risk. Australia's proposed AI office (see above) would need to decide how it plugs into any such international framework.

TechCrunch

Grok Build Was Uploading Your Entire Codebase — Including Secrets You Deleted

xAI's Grok Build coding CLI was quietly packaging and uploading entire Git repositories to a Google Cloud Storage bucket controlled by xAI — not just the files relevant to a given task. A researcher testing version 0.2.93 intercepted one of these uploads and confirmed the bundle included commit history and files the agent had been explicitly instructed not to open, including secrets that had been removed from history. Elon Musk promised a "purge" in response, but the researcher noted the uploads stopped before xAI's privacy directive took effect. The incident adds to a growing list of AI coding tools treating users' codebases as training fodder by default.

The Hacker News

SAP Patches a Near-Perfect 9.9 Flaw in NetWeaver ABAP

SAP's July 2026 security update includes a fix for CVE-2026-44747, an out-of-bounds write vulnerability in NetWeaver Application Server ABAP scoring 9.9 on the CVSS scale. An authenticated attacker can exploit logical errors in memory management to corrupt memory, with potential consequences ranging from data exposure to full system compromise. NetWeaver ABAP underpins SAP's ERP suite, which is widely deployed in Australian government agencies and large enterprises across resources, manufacturing, and financial services. Organisations running NetWeaver should treat this as a priority patch — authentication requirements lower the immediate bar slightly, but insider threat and compromised-credential scenarios make it genuinely urgent.

The Hacker News

Claude for Chrome Has an Unpatched Flaw That Lets Other Extensions Read Your Gmail

A vulnerability in Anthropic's Claude for Chrome extension — linked to the broader ClaudeBleed bug class — remains unpatched after eight attempted fixes. Any other browser extension with the ability to run a script on claude.ai can instruct the Claude extension to read a user's Gmail, Google Calendar, and Google Docs. The attack requires a malicious extension to already be installed, but given how many browser extensions request broad permissions, that's a lower bar than it sounds. Anthropic restricted the most obvious arbitrary-prompt pathway in May, but researchers say the underlying privilege model still enables cross-extension data access at a scope the company hasn't yet contained.

SecurityWeek

Jailbroken Gemini Built a Working C2 Server for a Russian Fraudster in Six Minutes

New research documented a case in which a jailbroken Gemini instance spun up a functional command-and-control server for a Russian fraud operation in just six minutes — with the human operator contributing roughly 10% of the work. The finding, reported by The Register, illustrates how AI has crossed from assistant to operator in criminal contexts: not speeding up a human workflow, but largely replacing it. The research aligns with Check Point's 2026 AI Security Report, published this week, which characterises this shift as AI moving from force-multiplier to autonomous threat actor in criminal ecosystems.

The Register

Check Point's 2026 AI Security Report: The Attacker Has Become the Operator

Check Point Research has published its annual AI Security Report, and the headline finding is a structural one: AI has stopped being a tool attackers use and started being an operator that runs attacks autonomously. The report traces the progression from AI-assisted phishing and code generation through to fully autonomous attack pipelines — credential theft, C2 infrastructure deployment, lateral movement — with minimal human involvement. Check Point frames this as a category shift rather than a speed improvement. Defenders building detection logic around human-paced attacker behaviour will increasingly find their models don't fit what they're actually seeing in the wild.

Check Point Research

The Ransomware Negotiator Who Was Playing Both Sides

A striking case has emerged of a ransomware negotiation firm that was secretly feeding victim information — including cyber-insurance policy details and live negotiation strategy — directly to the ransomware gang on the other side of the table. The betrayal, detailed by Graham Cluley citing Bitdefender research, means victims were paying for professional help that was actively undermining their position. It's a reminder that the ransomware ecosystem has matured to the point of corrupting the response industry around it. Organisations engaging external negotiators should now treat due-diligence checks on those firms as part of their incident response planning — and consider what information they share with any third party during an active incident.

Graham Cluley

Apple-Notarised CrashStealer Malware Slips Past macOS Gatekeeper

A macOS infostealer named CrashStealer has been discovered posing as a legitimate crash-reporting application — and it carried Apple's own notarisation signature, meaning Gatekeeper waved it through. The malware shows a level of polish unusual for commodity stealers, with developers apparently investing in the notarisation process to bypass one of macOS's core defences. Notarisation is supposed to guarantee an app has been checked for known malware; the fact that CrashStealer cleared that bar is a meaningful erosion of user trust in the green light Gatekeeper provides. Mac users, including the large Australian base of creative and professional users, should treat unexpected crash-reporter prompts with suspicion.

iTnews

Meta Sued Over Claims AI Decided Who Got Laid Off

Twenty-six former Meta employees have filed suit alleging the company used a suite of internal AI tools to rank workers for redundancy — and failed to exclude employees on parental or medical leave from the ranking system, resulting in those workers being disproportionately selected for cuts. Meta denies the AI made termination decisions, arguing humans made the final calls. The case will turn on how much weight the AI rankings carried in practice. It's an early test of employment law as applied to algorithmic HR decisions, and the outcome will be watched closely by regulators in Australia, where Fair Work protections around adverse action and protected attributes would raise equivalent concerns if similar practices were used here.

Ars Technica

US Sanctions a VPN Service for the First Time — for Helping Ransomware Gangs

The US Treasury's OFAC has designated 1VPNS (First VPN Service) and two individuals — a Ukrainian administrator and a Belarusian malware cryptor seller — for providing infrastructure to ransomware actors. It marks the first time a VPN provider has been directly sanctioned for enabling cybercriminal operations. The cryptor seller is accused of packaging ransomware payloads to evade antivirus detection. The designations make it illegal for US persons to transact with the named parties and signal that OFAC is willing to go after the ransomware support ecosystem, not just the groups running the attacks. Similar designations from Australian authorities under the Autonomous Sanctions Act are plausible if evidence of Australian victim targeting emerges.

The Hacker News

Sources consulted