- Published on
The week AI hacking tools stopped being a warning.
- Authors

- Name
- AbnAsia.org
- @steven_n_t
Listen to this article · The week AI hacking tools stopped being a warning.
0:00 / 0:00
Watch this article · The week AI hacking tools stopped being a warning.
Three disclosures landed in seven days, and they are not separate stories. On 13 November the FBI published an alert about an open-source framework that turns a language model into an intrusion agent. The same week, a worm moved through the remote-support tools managed service providers use to reach their customers. And a campaign used 4,837 stolen GitHub accounts to hunt for credentials in build pipelines.
Different actors, different targets, one mechanism: the cost of running an intrusion fell far enough that it stopped asking for a skilled human.
What the FBI actually described
ARTEX is not a vulnerability. It is a workflow. An agent takes a target, runs reconnaissance, picks an exploit, escalates and pulls data — and its operator watches. The alert says some intrusions go from first contact to privileged access in under three hours, and that agents let people with modest skill run several intrusions at once.
The alert also names where they go: partner and supplier portals, mobile backends, legacy services, third-party integrations. In other words, the systems that sit outside the security team's daily attention because they belong to someone else's project.
One more detail matters commercially. Amazon Threat Intelligence reported the same day that groups are using these agents to steal AI service keys and cloud credentials — the stolen keys then pay for the inference the next intrusion needs. The attack funds itself.
The worm that walked in through a supplier
The second story is the one enterprises should read twice. Huntress counted 288 endpoints at 48 organisations on 11 November, then a second wave of 205 hosts at 41 organisations two days later. Microsoft described the path: access to a managed service provider's cloud-hosted ScreenConnect panel, then attacker-controlled clients pushed to customers. The opening was a third-party extension — versions 1.0.31 and 1.0.32 accepted a blank access key when none had been configured.
ConnectWise disabled five extensions across more than 400 partner environments. The lesson is not "patch ScreenConnect". It is that one forgotten plugin on a supplier's console is a route into every customer that console can reach.
Credentials were the target, not the prize
The third disclosure closes the loop. Aikido reported 4,837 hijacked GitHub accounts probing 61,764 repositories with pull requests titled "docs: dependency audit notes". Where the account had write access, the attacker edited a CI workflow or planted an npm install hook, ran a 1,500-line credential stealer, then deleted the run and its logs in under a minute.
And the AI supply chain is now part of the same surface. Cursor disclosed that a stolen key at its AI provider let an attacker copy historical agent traces containing unredacted credentials, then publish a malicious copy of a coding-assistant extension in its marketplace. It stayed up for about eleven hours. Up to 5,000 developers were exposed.
One key is stolen
An AI provider, a CI pipeline, a support console — the key is the entry, not the exploit.
The agent does the work
Reconnaissance, exploitation and lateral movement run as a workflow, at machine speed.
The trust boundary is the path
Supplier console, marketplace listing, build pipeline: each one is trusted by many customers.
The data leaves before anyone looks
Runs and logs deleted in under a minute; detection arrives after exfiltration.
What to do before it is your week
There is no silver bullet, and anyone selling one is describing a product they have not run against an adversary.
What works is narrower, and it is unglamorous. Run a multi-model testing harness across your infrastructure — one model finds what another misses. Patch your own code. For open source you cannot patch on your own schedule, buy virtual patching and mean it. Assume the partner portal and the legacy integration are found first, because the alert says that is where they go.
Then look at your SOC. A stack of dashboards over separate data lakes cannot correlate an event that unfolds in ninety minutes across six systems. The only shape that gives you a chance is a real-time data lake with ML detectors running across all of it, and AI doing the correlation — because the attacker is already running one.
Band-aid solutions that work on multiple data lakes are not a plan. They are a delay.
When it happens — and on this evidence it will — the question will not be whether your team is skilled. It will be whether they can see it in time.
Source: FBI FLASH alert on the ARTEX agentic framework (13 November 2026); Amazon Threat Intelligence; Huntress and Microsoft Threat Intelligence on SCKit; Aikido on the wfx-loot campaign; Cursor incident disclosure; BleepingComputer reporting on Barracuda and Chainalysis. Figures are as published by each source on the dates cited.