Here's the ugly one. After a January 1 selfie of Amani Al-Khatahtbeh in a black hijab and purple fur coat, replies asked Grok to strip the hijab and put her in lingerie. Grok did it, and the edited images reached tens of thousands of people. She filed a federal complaint Wednesday alleging intentional distribution of nonconsensual sexual images, design defect, negligence, and intentional infliction of emotional distress.
If you're building on generative image models, this is the part to read twice. The complaint's core move is calling Grok a defective product: it argues xAI could have blocked prompts that remove religious garments from real people, and that a model trained on large uncurated image datasets should refuse that request by design. That framing turns a content moderation failure into a product liability claim.
The scale is what makes this hard to wave away. A January report estimated Grok produced around 3 million photorealistic sexualized images on X in 11 days, with about 23,000 depicting children. A separate review of 500 images from January 6 to 9 found five percent contained these hijab-targeting edits. xAI limited the tools to paying subscribers, and it hasn't publicly responded to requests for comment.
Attackers are now running AI agents that scan, probe, and break into networks almost autonomously, and the most prominent framework is Artex. It's an open-source Go project that wires frontier or open-weight models to existing penetration testing tools, so a low-skill operator can type a plain-English goal and let multiple agents break it into tasks. Since late August, banks, government agencies, and companies have all been hit.
If you run infrastructure, this changes your threat model, not your attack surface. Artex doesn't invent new exploits. It just compresses the timeline. Attacks that used to take days now finish in hours, and Microsoft saw Chinese actors try to sign into Microsoft 365 tenants within a few hours of a supply chain compromise. The good news is that these tools are noisy, firing thousands of probes in rapid succession, so your existing monitoring can catch them if you're actually watching.
The numbers are worth knowing. Researchers have catalogued 45 AI attack frameworks so far. Vapor Panda, a China-linked group, used Artex against more than 200 organizations in October. Censys and Shadowserver found thousands of Artex instances exposed to the internet, some with no authentication at all. The original GitHub repo was pulled, but forks keep circulating.
What to watch: whether the safety guardrails in frontier models keep blocking these workflows, or whether open-weight models make them moot. Let's dive deeper…
In today's ContentBuffer update: