AI and IT News Recap: August 4, 2026: Alibaba's Qwen3.8-Max Rivals the Frontier on the Cheap, INC Ransomware Storms SonicWall VPNs, and Malware Walks Past Google's Passkeys

By Noah Smith, Founder and Principal Consultant, KeyChange Technologies · August 4, 2026

A quieter day on the model front turned into a loud one when Alibaba dropped a 2.4-trillion-parameter model that lands within striking distance of the Western frontier, while the security world spent the day cleaning up after ransomware crews, poisoned packages, and a set of attacks that walk straight past your passkeys.

📌 The AI and IT news at a glance

• Alibaba's Qwen3.8-Max rivals Claude Opus 5 on coding benchmarks at a fraction of the price, and it is about to go open source.
• INC Ransomware has become the dominant crew exploiting two SonicWall SMA 1000 VPN flaws to reach corporate networks.
• Malware can now bypass Google's passkeys on Windows without a fingerprint, PIN, or anything showing on your screen.
• N-able's N-central fix was incomplete, and attackers are taking over the servers MSPs use to manage your endpoints.
• "FaceHugger" flaws in Hugging Face Diffusers let a poisoned AI model run code on the machine that loads it.
• 18 malicious npm packages are pushing a cross-platform remote access trojan at developers using Alibaba tools.
• Cato Networks shipped Agentic Threat Prevention, using autonomous agents to block attack paths before they are used.

🔝 Top story: Alibaba's Qwen3.8-Max crashes the frontier party

Alibaba used August 3 to unveil Qwen3.8-Max, the most capable model it has ever shipped and roughly seven times larger than the Qwen3.5 model it released back in February. The headline number is 2.4 trillion total parameters, but the model only activates about 95 billion of them to answer any given query, which is how a model this big stays affordable to run. It handles prompts up to 1 million tokens, enough to chew through more than 200 pages of text or about 100 hours of footage in a single request, and it can return answers up to 131,000 tokens long.

The reason this matters beyond the spec sheet is where it landed on the leaderboards. Qwen3.8-Max scored 1,668 on the Frontend Code Arena benchmark, putting it just 37 points behind the strongest configuration of Anthropic's Claude Opus 5 and ahead of more than a dozen other frontier models. Alibaba says the model finished a 16-day coding project with no human input and completed a chip-design optimization task running over 500 steps. It is available now through Alibaba Cloud, and the company plans to open-source it next week alongside a smaller, more hardware-friendly Qwen3.8-27B version, all at pricing that sits well below the Western frontier labs.

In short: Alibaba launched Qwen3.8-Max, a 2.4-trillion-parameter model that nearly matches Claude Opus 5 on a key coding benchmark and will be open-sourced next week.

What it means for your business: The gap between the expensive frontier models and cheap, open, self-hostable ones keeps shrinking, which gives you real leverage on price and more freedom to run capable AI on your own terms.

My take: The number that should stick with you is 37 points behind Opus 5, not 2.4 trillion parameters. A model you can soon download and run yourself getting that close to the best commercial systems changes the math on vendor lock-in. It also comes from China, so anyone in a regulated or government-adjacent industry should factor data governance and provenance into the decision before piloting it. Cheaper is only better if it fits your compliance posture.

Source: SiliconANGLE

🛡️ IT and security

INC Ransomware becomes the dominant crew hitting SonicWall SMA 1000 VPNs

Security firm Resecurity reported over the weekend that INC Ransomware has emerged as the dominant threat actor exploiting a pair of recently patched flaws, CVE-2026-15409 and CVE-2026-15410, in SonicWall's Secure Mobile Access (SMA) 1000 series VPN appliances. Chained together, the two bugs let an unauthenticated attacker run commands and take over the device. SonicWall shipped fixes in mid-July, but the flaws had been exploited as zero-days since at least June 22, and the group has now claimed 885 victims on its leak site, with the most recent added on August 2.

What makes this one nasty is what the attackers grab once they are in. Rapid7 says the intrusions pulled high-value credentials, active session databases, and Time-Based One-Time Password (TOTP) multi-factor seeds, the ingredients needed for long-term, persistent access and lateral movement deeper into the network. Some victims have even reported harassing phone calls and emails from people claiming to be hackers, a pressure tactic to force payment.

In short: INC Ransomware is chaining two patched SonicWall SMA 1000 flaws to steal credentials and MFA seeds and burrow into corporate networks.

What it means for your business: If you run a SonicWall SMA 1000, patching alone is not enough now, because stolen session tokens and MFA seeds mean attackers may already hold keys that survive the update.

My take: This is the classic "patch is necessary but not sufficient" trap. Once TOTP seeds and session databases are out the door, updating the box does nothing to lock out an attacker who already copied the keys. If you have one of these appliances, treat it as compromised until proven otherwise: rotate credentials, reset MFA enrollments, and go hunt through your logs for the lateral movement that usually follows.

Source: The Hacker News

Malware can walk straight past Google's passkeys on Windows

Passkeys are supposed to be the thing that finally kills the password, so it is worth paying attention when researchers show malware defeating them quietly. Palo Alto's Unit 42 detailed three attack paths against Chrome's Google Password Manager cloud authenticator, nicknamed Pass-ta-key, Silver Pass-ta-key, and Golden Pass-ta-key. Running as an ordinary user on a Windows machine, the malware can sign into a victim's passkey-protected accounts with no fingerprint, no PIN, and nothing appearing on the victim's screen.

Importantly, none of this breaks the underlying cryptography. The attacks target the code around the passkey: how Chrome stores its device keys, how it re-enrolls a device after that state disappears, and whether the website you are logging into even bothers to confirm a human was verified. The strongest path can extract the 32-byte Security Domain Secret used to decrypt a user's synced passkey private keys, effectively handing the attacker the whole set.

In short: Unit 42 showed that ordinary Windows malware can silently authenticate to Google passkey-protected accounts and even extract the secret that decrypts synced passkeys.

What it means for your business: Passkeys are still a big upgrade over passwords, but an endpoint already infected with malware is not safe, so device security remains the foundation everything else sits on.

My take: Do not read this as "passkeys are broken," because they are still far better than reused passwords. Read it as a reminder that no login method rescues a machine that is already owned. The practical takeaway is boring but true: keep endpoint protection current and treat compromised laptops as a full-account incident, not just a "reimage and move on" chore.

Source: The Hacker News

N-able's incomplete N-central fix lets attackers own the servers that manage your IT

N-able warned that attackers exploited an authentication bypass, CVE-2026-18577, in its N-central platform to gain remote administrative access, and that its first fix did not fully close the hole. N-central is the remote monitoring and management tool that managed service providers and internal IT teams use to administer customer endpoints, so a compromise here is a compromise of everything downstream. The company shipped build 2026.3.1.7 on August 2 as the first version that is actually unaffected.

Once inside an N-central server, the attackers used its built-in Take Control feature to reach managed endpoints, then registered Cloudflare tunnels as services on those devices. Because the tunnels connect outbound to Cloudflare's edge, they need no inbound firewall rule or open port, and running them as services means they survive a reboot and keep working even after the original path through the N-central server is cut off. Nothing suggests Cloudflare itself was compromised; the attackers simply abused a legitimate service.

In short: A patched-but-then-repatched auth bypass in N-able N-central let attackers reach managed endpoints and plant persistent Cloudflare tunnels.

What it means for your business: If your IT is handled by an MSP, ask whether they run N-central and whether they are on build 2026.3.1.7, because a breach of that tool reaches straight into your machines.

My take: Management platforms are the softest, highest-value target in any IT stack, because one server owns hundreds of endpoints. The persistence trick here is the part to remember: revoking the obvious access route did not evict the attacker, because the Cloudflare tunnels kept humming. Patching is step one; step two is actually hunting for the tunnels and services that outlive the fix.

Source: The Hacker News

"FaceHugger" flaws turn a Hugging Face AI model into a code-execution trap

Researchers at Zafran Labs disclosed three high-severity flaws, collectively named FaceHugger, in Hugging Face's Diffusers library, the widely used Python package for running image and media generation models. The flaws let a crafted model repository quietly execute arbitrary code on any machine that loads it, and they do so by bypassing trust_remote_code, the very safeguard designed to stop unreviewed code from running in custom pipelines.

This matters because Hugging Face has become the GitHub of the AI era, and libraries like Diffusers are baked deep into production pipelines, CI/CD systems, and container images. A developer who pulls a poisoned model to try it out could be handing an attacker code execution inside the build system, not just on a laptop.

In short: Three flaws in Hugging Face's Diffusers library, dubbed FaceHugger, let a malicious model repository run code on the machine that loads it by bypassing a key safety control.

What it means for your business: Downloaded AI models are now executable content, so treat a model from a public hub with the same caution you would give an unknown software package.

My take: The mental model shift here is important. Teams still tend to think of an AI model as data, a big file of weights, when in practice loading one can run code. If your developers are pulling models off public hubs, that is now part of your software supply chain and deserves the same review, pinning, and scanning you already apply to your other dependencies.

Source: The Hacker News

18 malicious npm packages push a cross-platform RAT at developers

Researchers found a fresh batch of 18 malicious npm packages aimed at developers using Alibaba's tools, delivering a cross-platform remote access trojan in a targeted supply chain attack focused on Chinese-speaking environments. The standout is "lib-mtop," an unscoped package that borrows the exact name of a private Alibaba package that normally lives under the protected "@ali" scope, a classic bit of confusion designed to trip up an autocomplete or a copy-paste.

The package sat harmless since November 2023 before three new versions were quietly uploaded this spring, each adding a loader that uses curl to fetch and run a remote JavaScript payload. Whether the maintainer account was hijacked or went rogue is not yet clear, but the pattern is familiar: a trusted-looking name, a long dormant period, then a sudden weaponized update.

In short: A set of 18 malicious npm packages impersonating Alibaba tooling is delivering a remote access trojan through a targeted supply chain attack.

What it means for your business: Dependency names that look almost right are a real attack vector, so lock your builds to known-good package versions rather than always pulling the latest.

My take: Naming attacks like this work because they exploit trust and muscle memory, not any technical flaw. The defense is unglamorous and effective: pin your dependencies, review what gets added, and be suspicious of a package that lay dormant for two years and then suddenly grew a network loader. Speed of adoption is the enemy of supply chain safety.

Source: The Hacker News

🛠️ New tooling and business AI

Cato Networks turns autonomous agents loose on stopping AI-assisted attacks

Cato Networks announced general availability of Agentic Threat Prevention, a capability that uses autonomous agents to predict the route an attacker is likely to take through a specific customer's network and then generate and enforce protections for that environment before the attack can advance. It stitches network and security telemetry into a single view and pushes the resulting protections out globally through every Cato point of presence, with no service chaining or enforcement gaps.

Cato says its internal agentic red-team lab has run more than 500 autonomous attack simulations since mid-May and that the platform consistently blocked the attackers before they reached their goals. The launch builds on the company's earlier Agentic CVE Mitigation feature, which it says can assess a newly disclosed vulnerability and apply protection in as little as 45 minutes.

In short: Cato Networks launched Agentic Threat Prevention, which uses autonomous agents to predict and block likely attack paths before they are exploited.

What it means for your business: As attackers start using AI to move faster, defense tooling that pre-positions protections automatically is becoming a practical way for lean teams to keep pace.

My take: The honest question with any "agentic" security product is whether it genuinely reduces work or just relocates it into tuning and false-positive triage. The 500-simulation figure is a reasonable signal, but it is the vendor's own lab. If a tool like this is on your radar, ask for a trial against your real environment and measure whether it actually shortens the window between a new CVE and a working defense. That 45-minute claim is the number worth testing.

Source: SiliconANGLE

That is your AI and IT news for August 4, 2026. For yesterday's rundown, see the August 3 recap.

Written by the KeyChange Technologies editorial team. KeyChange Technologies is an IT and AI services firm that helps small and mid-size businesses adopt modern technology securely. We track the AI and security landscape daily so business owners do not have to.