Daily brief at 7am Melbourne. Unsubscribe any time.

Sunday 18 October 2026

One Stolen Key, 173 Front Doors: The ConnectWise M365 Attack Is an MSP's Worst Nightmare

A compromised ConnectWise credential unlocked 173 Microsoft 365 tenants simultaneously — a textbook MSP supply chain attack that should terrify anyone relying on a managed service provider.

Lead story

One Stolen Key, 173 Front Doors: The ConnectWise M365 Attack Is an MSP's Worst Nightmare

Imagine handing your building's master key to a security contractor, and then that contractor gets pickpocketed. That's essentially what happened here. Threat actors obtained a single compromised Microsoft 365 credential belonging to ConnectWise Cloud Backup — the platform formerly known as SkyKick — and used it to waltz into 173 separate M365 customer environments without triggering a single password prompt.

ConnectWise Cloud Backup is a backup and recovery tool built specifically for managed service providers (MSPs). MSPs are, by design, trusted to hold elevated access to dozens or hundreds of client tenants at once. That's the value proposition. It's also the attack surface.

What actually happened. The attackers didn't exploit a software vulnerability or write a single line of malware. They simply authenticated as ConnectWise. Once inside the ConnectWise service account, they had delegated administrative access to every tenant that had granted the backup service permission to touch their M365 data. At last count, 173 tenants had been accessed — and that number may still be moving.

The specific nature of the access is what makes this alarming. MSP tooling typically holds broad permissions: mailbox access, SharePoint, Teams, OneDrive. Whoever was in there could read email, exfiltrate files, or quietly establish persistent access via new accounts or OAuth tokens — all without the individual customer organisations ever seeing a suspicious login on their own tenancy.

Why MSPs keep being the target. This isn't the first MSP-as-pivot-point incident, and it won't be the last. The economics are compelling for attackers: compromise one supplier, inherit access to dozens of downstream targets. We saw the same playbook in the SolarWinds campaign, the Kaseya VSA ransomware attack, and a string of smaller incidents since. Each time, the lesson is the same — third-party trust relationships are a force multiplier, for defenders and attackers alike.

What to watch for. ConnectWise has not yet disclosed whether affected tenants have been notified, whether the initial credential was phished, bought on a dark web market, or obtained some other way. The "how" matters for defenders trying to understand their own exposure.

For Australian organisations, the implications are direct. The MSP model is the backbone of IT delivery for a vast swath of Australian small and medium businesses, and the ACSC's Essential Eight specifically calls out third-party access management as a control area. Under the SOCI Act, critical infrastructure entities using MSPs are required to assess and manage supply chain risk — incidents like this are exactly the scenario those obligations were written for.

What to do right now. If your organisation uses ConnectWise Cloud Backup — or any MSP tooling with delegated M365 access — this is the moment to pull your unified audit logs and look for unexpected sign-ins, new OAuth app registrations, or unfamiliar admin accounts. Don't wait for ConnectWise to notify you. Assume breach, verify clean.

The deeper lesson here is a familiar one: federated trust at scale is a liability until it's actively managed. One credential, 173 consequences.

Also today

Candiru's Spyware Hitched Rides on Perfectly Normal-Looking Ads

Israeli surveillance vendor Candiru found a clever new delivery mechanism: a legitimate-looking ad firm called Discover-Tech. According to new research, Discover-Tech's ad infrastructure was quietly being used to fingerprint devices at scale, allowing Candiru clients to identify high-value targets and then silently deliver zero-click spyware — no interaction required from the victim. It's a troubling evolution because ad networks are ubiquitous and largely trusted. The targeting layer means victims are selected surgically rather than sprayed broadly, making detection harder and the potential for political targeting much higher. The findings underscore why commercial spyware vendors remain one of the most underappreciated threats to civil society.

SecurityWeek ↗

Malicious npm Package 'imgdecode' Smuggles a Shellcode Loader Inside a Windows Native Addon

A package called 'imgdecode', uploaded to npm on 30 September, has been identified as part of a broader software supply chain campaign researchers are calling MALFEX. The trick here is subtle: the malicious payload is hidden inside a Windows native addon — a compiled binary component that npm packages can include to run platform-specific code. This approach sidesteps many text-based static analysis tools that scan JavaScript files but don't deeply inspect compiled binaries. The package had been downloaded nearly 980 times before disclosure. It remains live on the npm registry, which is its own problem. Developers relying on automated dependency scanning without binary analysis are flying partially blind against this class of attack.

The Hacker News ↗

Cybercriminals Are Now Locking People Out of Their Own Office Buildings

Data theft was so last decade. A new wave of financially motivated attackers is going after building management systems — the software controlling door access, HVAC, and physical security in commercial properties. Researchers have documented incidents where criminals compromised building controllers and threatened to lock employees out, disable climate systems, or trigger false fire alerts unless a ransom was paid. The attack surface is real: building management systems are increasingly IP-connected, frequently run outdated firmware, and are almost never included in a typical IT security audit. For facilities managers and physical security teams who've never had to think about threat actors, the learning curve is steep and the window is closing. This is OT security's next front.

The Register ↗

Encryption Is Getting Squeezed From Every Direction at Once

Canada's proposed surveillance legislation is the latest addition to a growing global list of threats to end-to-end encryption. Wired's analysis maps out how governments from the UK to the EU to Australia have each, in their own way, pushed for lawful access provisions that would require tech companies to build backdoors or break encryption on demand. The problem is structural: a backdoor for law enforcement is a backdoor for everyone. Australia's Assistance and Access Act already sits in this category, and the ACSC has previously warned about the risks of weakened encryption to national cyber resilience. When multiple democratic governments move in the same direction simultaneously, platform providers face enormous pressure to comply — and the global internet suffers.

WIRED Security ↗

US Federal Agencies Release Their Most Detailed AI Sandbox Security Guidance Yet

A joint advisory from US federal agencies has laid out voluntary but unusually detailed cybersecurity controls for organisations testing frontier AI models in sandbox environments. The guidance covers isolation requirements, monitoring controls, and recommendations for independent third-party testing before deployment. While it carries no enforcement weight, it's significant as a benchmark — the most specific the US government has been about what responsible AI security testing should look like in practice. For Australian organisations developing or deploying frontier AI, the document is worth reading alongside the ASD's own AI security guidance. The voluntary framing mirrors Australia's current approach, though calls for mandatory standards are growing louder on both sides of the Pacific.

CyberScoop ↗

Replit Drops Anthropic as Default, Switches to OpenAI

Replit, the browser-based coding environment popular with developers and students, has quietly shifted its default AI model provider from Anthropic to OpenAI. CEO Amjad Masad confirmed the move over the weekend, noting that Claude will still be available on the platform for users who want it — it's just no longer the default. The decision is notable because Replit was one of Anthropic's more prominent product partnerships and a visible use case for Claude in developer tooling. It's another data point in the ongoing contest for AI model market share, where default placement is everything — most users never change what comes out of the box. For Anthropic, losing a flagship developer platform default is a meaningful signal even if the revenue impact is modest.

TechCrunch ↗

Websites Are Starting to Block AI Agents — and It's Getting Messy

As AI agents from OpenAI, Meta, Google, xAI and others start doing real things on the web — booking reservations, checking inventory, scraping product data — a growing number of websites are simply blocking them. Users of services like Meta's Muse and OpenAI's Dots have reported their agents hitting walls on major platforms including Yelp. The tension is understandable: AI agents consume server resources, bypass ad revenue, and often act in ways that violate terms of service. But the blocking is inconsistent and sometimes catches legitimate users. This is shaping up to be the next big negotiation between platforms and AI companies — the same fight over scraping for training data, but now in real time and at operational scale.

The Verge ↗

Meta Quietly Removes Terms That Would Have Let It Scrape News Sites for AI Training

Meta has updated its commercial terms of service to remove language that, as written, would have allowed it to scrape content from any website — including news publishers — to train its AI models and improve its products. The removal came after backlash from publishers and legal observers who flagged the clause as unusually broad. Meta hasn't explained why the language appeared in the first place or whether any scraping occurred under it. Australian news publishers, who fought a long battle over the News Media Bargaining Code, will be watching closely — the ACCC has signalled ongoing scrutiny of AI companies' data practices, and this incident shows why precise terms-of-service language is not a trivial matter when AI training is involved.

The Verge ↗

TikTok Permanently Bans Account Promoting Musk Documentary

A four-hour investigative documentary about Elon Musk by filmmaker Alex Gibney hit a wall on TikTok this week, with the platform permanently banning the film's promotional advertising account just before the documentary's wider release. TikTok gave no public explanation. The timing has raised eyebrows given Musk's ongoing business entanglements with China-adjacent entities and TikTok's parent company ByteDance's Beijing connections. Whether this was algorithmic, legal, or something more deliberate is unclear — TikTok isn't saying. What's clear is that a major streaming platform's promotional channel being silenced on the world's largest short-video app, right before launch, is a significant distribution setback and a fresh example of platform gatekeeping in action.

TechCrunch ↗

SpaceX Is Betting It Can Actually Replace the Major US Carriers

SpaceX has agreed to acquire a significant block of low-band wireless spectrum — the kind that travels long distances and penetrates buildings — with the ambition of building Starlink into a full mobile carrier capable of competing with AT&T, T-Mobile, and Verizon. Analyst Walter Piecyk describes the plan as expensive and genuinely risky, but not impossible. Low-band spectrum is the foundation of rural and suburban coverage, and SpaceX's satellite infrastructure gives it a different cost structure to traditional tower-based carriers. For Australia, where rural connectivity remains a persistent problem and Starlink has already signed government contracts for remote areas, a vertically integrated SpaceX mobile offering could reshape the competitive dynamics of outback connectivity.

The Verge ↗

Anthropic's Workforce Is Maturing Fast — and That Changes How It Operates

Anthropic has disclosed that the makeup of its workforce has shifted substantially over the past year: fewer junior employees, more experienced hires, and — perhaps as a consequence — more time spent in meetings. The company didn't frame this as a problem, but the pattern is familiar in fast-scaling tech organisations. As headcount grows and seniority rises, coordination costs increase, decision-making slows, and the scrappy research-lab culture that drove early breakthroughs can quietly erode. For a company that's staked its identity on safety-first AI research, the question is whether a more corporate internal culture changes the quality or character of that work. It's a softer signal than a product launch, but worth watching as Anthropic competes with OpenAI for talent and mindshare.

TechCrunch ↗

Sources consulted