
Table of Contents
THE SIGNAL
On Friday the FBI issued a flash alert on ARTEX, the open-source agentic attack framework behind the WorkOS breach I covered last week. The Record reports that the bureau sees criminals and nation-states using it to run intrusions from reconnaissance through exfiltration, in some cases reaching privileged access in under three hours. The original GitHub repository has been removed, and it hasn't mattered. The tool survives in forks, archived copies and a commercial version, and it will drive whichever model you plug in. In the attacker instances the FBI analysed, DeepSeek models were the most common. Victims include financial institutions in South Korea and Japan, and Microsoft says Chinese state-backed actors used it in the supply chain attack on ConnectWise.
That is the loud part. In the same seven days, three more agent frameworks turned up on real victims. Microsoft Threat Intelligence describes Storm-3358, a financially motivated cluster based in Cambodia, pointing ARTEX at financial organisations since October: mostly in Southeast Asia, plus a smaller number of large global firms. Arctic Wolf found AntiFox, a Chinese-language ARTEX fork, behind intrusions on internet-facing IIS servers and AWS environments, where at least some calls against a production secret succeeded. Acronis traced BOSSAI, a 12-agent setup, to one person in Hualien City, Taiwan. And CloudSEK found an exposed server showing Silent Lynx, a Kazakhstan-linked espionage group, steering an open-source agent called Hermes from a Telegram bot while it stole thousands of Tajik Defence Ministry documents.
The part to read closely is how they got in, and where they failed.
BOSSAI is the best documented. Its agents scanned hundreds of organisations in at least 11 countries and got data or access at about ten. The wins were Taipei's 119 emergency dispatch records pulled from a public-facing API, a fire department platform serving 2018 patient data without a login, a database admin panel still on default root credentials, and SQL injection at a lending firm and three Thai universities. Acronis says the agents failed wherever a login, a valid session or endpoint filtering was enforced. The operator also had to run "abliterated" local copies of DeepSeek, GLM and Qwen, because commercial models including Grok refused the attack requests.
The other cases rhyme. Arctic Wolf watched AntiFox make hundreds of authenticated AWS calls in about 30 minutes, and most were denied. Silent Lynx reached Tajikistan's national document platform with one railway employee's credentials. Microsoft reports that honeypots, tarpits and fake wins reduced ARTEX traffic against real assets in its testing.
I won't oversell this. Arctic Wolf doesn't know how the IIS server was first compromised, and Storm-3358's agents did exploit file upload and SSRF flaws. But nothing reported this week shows an agent beating a control that was working. The agents have made trying every door nearly free. They haven't shown they can open a locked one.
Sources: The Record (FBI flash), Microsoft Threat Intelligence, BleepingComputer (AntiFox), The Hacker News (AntiFox), The Hacker News (BOSSAI), The Record (Silent Lynx).
THE MAP

The agent multiplies your exposure. Your locks still hold. Hundreds of organisations were scanned and about ten were breached, each through something already open. The controls that stopped the rest were ordinary ones: a login, a session check, a filter. Deception adds a fourth, because an agent that chases every win will also chase a fake one.
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
AROUND THE PERIMETER
Check Point Mobile Access: attackers chained the PHP2Shell flaws (CVE-2026-109618, CVE-2026-109619) with a zero-day in Gaia OS (CVE-2026-123039) to get root on the gateway, then scanned for internal ESXi hosts and ran a wiper. Check Point confirms roughly two dozen customers exploited since late October. Apply the hotfix, and confirm your ESXi recovery works without the hosts being wiped.
ScreenConnect RMM+ and N-able N-central: two MSP management tools fell in one week. A third-party ScreenConnect extension accepted a blank access key, giving an attacker 288 endpoints at 48 organisations in about 90 minutes. N-central took more than 30 published zero-days, shipped a hotfix, and then saw 21 servers compromised, 19 of them already running it. If an MSP touches your estate, ask this week which of the two it runs.
Qubole: someone with credentials for Qubole's own AWS account assumed cross-account roles in customer environments and harvested Secrets Manager, Parameter Store and Lambda variables. More than a dozen organisations were hit, mostly through roles left in place after they stopped using the product. A vendor you have offboarded can still hold a working key.
wfx-loot on GitHub: 4,837 hijacked accounts were used to probe 61,764 repositories and push a CI secret stealer through pull requests titled as dependency audit notes. Secrets were stolen from at least 1,835 targets, Fortune 100 companies among them.
Chrome, CVE-2026-125350: a use-after-free in the XML component, exploited in the wild, and the eighth Chrome zero-day this year. Push 157.0.8103.50 and check your other Chromium-based browsers.
CALM THE NOISE
Patch Tuesday produced the week's biggest number, and nobody agrees on what it was. BleepingComputer reported a record 1,235 flaws, The Hacker News 1,259, The Register 1,094 and Krebs 1,093. A count that moves by 166 depending on who is counting is a poor input to a threat model.
The useful numbers are smaller. One of those flaws is being exploited: CVE-2026-124194, an Outlook spoofing bug scored 6.5 that can leak a user's NTLMv2 hash when they open a crafted message file. Two unexploited ones belong next to it because they need no click, Windows Search through Exchange attachment indexing (CVE-2026-116928) and Windows LDAP (CVE-2026-119958), both scored 9.8. So that's three patches this week and everything else on the normal cycle. Krebs notes that Microsoft credits AI for finding many of these bugs, so the monthly totals will keep growing. Every one of them is a bug the defending side found and closed.
The other loud story was Microsoft's backdoor research. Rewriting seven of more than 12,000 neurons in one layer of an open-weight model made it emit an attacker's command in at least 98.8% of triggered sessions while still passing ordinary benchmarks. A companion study hid 90 poisoned examples in a 9,000-conversation fine-tuning set and reached 78% attack success. It matters if your teams download open weights or fine-tune. But Microsoft says it has seen no adversary use the neuron method, and its detector flagged 9.7% of clean checkpoints. Put model provenance on the roadmap. This week's victims fell to default passwords.

