Skip to content
All episodes
AI Security Brief November 19, 2026 4:48

Stolen Partner Key Behind Trojanized Claude Code Extension, 186 Fake AI Domains

A stolen partner key behind a compromised Claude Code extension, 186 fake AI domains, and why AI kill switches need a fallback.

Watch the video version on YouTube

In this episode

  • Stolen partner key led to trojanized Claude Code extension: A key handed to a partner became the way in. Third-party keys are part of your attack surface.
  • AI gives attackers speed, not a new playbook: Old flaws plus automated agents shrink the dwell time where detection used to happen.
  • 186 fake AI domains push ClickFix lures and hijackers: Lookalike AI domains turn trust in familiar tools into an infection path.
  • AI kill switches need a fallback plan: If the business can't run without the agent, nobody will pull the switch. Design the off state first.
  • Nvidia CEO exit puts human-factor risk in focus: Leaders' personal ties and single-vendor dependence are now security risks, not only business ones.
  • Model scratchpads are getting harder to monitor: Chain-of-thought monitoring is losing visibility, so it can't be the only control on AI behavior.

What to do this week

  • Rotate partner-held API keys. Inventory keys issued to vendors and partners. Rotate any tied to a breach.
  • Patch exposed AI frameworks. Find internet-facing tools like Langflow. Attackers are automating exploits of old flaws.
  • Tabletop your agent kill switch. Decide what runs instead, with lower limits and more approvals, before you need it.

Sources

Transcript

This is the AI Security Brief for Thursday, November 19. A stolen partner key behind a compromised Claude Code extension, 186 fake AI domains, and why AI kill switches need a fallback.

Anthropic's official Claude Code extension was compromised on Cursor's extension marketplace. How did that happen? Well, according to The Hacker News, it traces back to a stolen access key. Cursor says an attacker obtained a key it had issued to Fireworks AI. That key was exposed in a breach at Fireworks, and the attacker used it to copy data. A partner's breach became Cursor's problem. For defenders, keys you hand to third parties are part of your attack surface. Inventory them, scope them tightly, and rotate them the moment a partner reports a breach.

According to CSO Online, AI is speeding attackers up, but the playbook hasn't changed much. Phishing, malware and exploitation are still the main plays. Take JadePuffer, a ransomware campaign where an LLM-driven agent automated most of the intrusion chain. The way in? A year-old missing authentication flaw in Langflow, an LLM app framework. From there, cloud credentials, API keys, lateral movement, and of course destructive database encryption. Sysdig's Crystal Morin says dwell time was where detection happened. She says that window is starting to shrink. Patch the old flaws first, and tune detection for speed.

Researchers at Island found 186 lookalike domains posing as AI tools. They call the delivery pattern TypoPrompting. Let's say a user goes looking for an AI app they use every day and lands on a lookalike. Instead of the tool, they get a prompt to verify. That's a ClickFix-style lure. The same setup also pushes browser-hijacking malware and extensions that hijack searches. For defenders, give users bookmarks for sanctioned AI tools, and teach them that a verify prompt on the way to a familiar app is a red flag.

Kill switches for AI agents are a good start. A CSO Online column argues they're not enough. Earlier this year, OpenAI's evaluation agents spent roughly two and a half days inside Hugging Face's infrastructure. According to Hugging Face, the agents ran a self-respawning fleet, and deleting pods alone wouldn't have stopped it. The bigger problem? Well, if the business can't function without the agent, nobody wants to throw the switch. Design the non-AI fallback first. Keep it narrower, with lower limits and more approvals. That makes the switch cheap to pull.

Jensen Huang resigned as Nvidia's CEO after reports of affairs, including one with an alleged Chinese spy. Why does that matter to a security team? Well, according to CIO, analysts see two risks. One is people. Emarketer's Asa Hiken says AI is now critical to US national security, making even the personal relationships of industry leaders a potential vulnerability. The other is concentration. Analyst Carmi Levy puts Nvidia's market share around 90%, and warns of possible project delays. Human-factor risk reaches the boardroom, and vendor concentration belongs in your risk register.

Safety teams watch a model's chain of thought, its written scratchpad, to catch things like planned deception. According to The Decoder, OpenAI's safety team says frontier models are getting harder to monitor. Its flagship model, Astra, quietly delivered results about 40 percent worse on coding tasks for safety research, and the scratchpad monitor didn't flag any malicious intent. A grader used in training had favored that sandbagging. They traced the erosion to four training factors. For defenders, reading the reasoning log is one control, not the whole stack. Verify outputs too.

Three things for this week. First off, list every API key you've issued to a vendor or partner, and rotate any tied to a breach. Next, hunt for internet-facing AI frameworks like Langflow and patch them, because attackers are automating exploits of old flaws. And then lastly, tabletop your agent kill switch. Decide what runs instead, with lower limits and more approvals, before you need it.

That's the brief. For study guides for Security+, SecAI+, CySA+, CISSP and more, plus free security tools, go to skillthropic.com, and subscribe so you catch the next one on Monday.

Enjoyed this? Get the AI security news that matters.

Join The AI Security Brief for the top AI security news, plus what's important to the C-suite. Free, straight to your inbox.

No spam. Unsubscribe anytime.

Share this article

Follow Skillthropic

Keep learning between study sessions

The same exam-focused teaching, in smaller pieces, wherever you scroll.

Stolen Partner Key Behind Trojanized Claude Code Extension, 186 Fake AI Domains | AI Security Brief