Skip to content
All episodes
AI Security Brief November 16, 2026 4:44

Backdoored Claude Code Extension, AI Ransomware Boom and Hijacked Agents

A backdoored Claude Code extension, ransomware crews leaning on AI, and a hidden line of text that rewires an invoice agent.

Watch the video version on YouTube

In this episode

  • Stolen key led to a trojanized Claude Code extension: Unredacted credentials in AI agent traces turned a vendor's leaked key into a supply chain attack.
  • AI lowers the bar in a ransomware boom: Attackers use AI for phishing and negotiation, letting inexperienced operators launch their own groups.
  • One hidden line turns an AI agent into payment fraud: Agents that read untrusted input and can change records are a fraud path most teams never review.
  • OpenAI launches Daybreak tools for AI pen testing: AI agents that find and fix flaws can speed up remediation, as long as a human still signs off.
  • AI chatbots point shoppers to fake retail sites: Users trust chatbot links, and impostor or re-registered domains are slipping into the answers.
  • Australia moves to mandate AI safety audits: Frontier labs would face audits and immediate incident reporting, a sign of where AI rules are heading.

What to do this week

  • Treat agent logs as secrets. Scrub credentials from AI traces and rotate any keys shared with AI vendors.
  • Inventory and review your AI agents. Flag any agent with untrusted input, sensitive data and the power to act.
  • Warn users before the holiday rush. Remind staff to verify any shopping link a chatbot gives them before paying.

Sources

Transcript

This is the AI Security Brief for Monday, November 16. A backdoored Claude Code extension, ransomware crews leaning on AI, and a hidden line of text that rewires an invoice agent.

Cursor has confirmed a breach that put a malicious copy of Anthropic's Claude Code extension in its marketplace. According to BleepingComputer, it started with an access key exposed at Fireworks AI, a third-party provider. Attackers used it to copy historical AI coding-agent traces, and some of those held unredacted API credentials. On October 29, the poisoned extension sat in the marketplace for about 11 hours. Up to 5,000 developers may have pulled it, and it tried to plant persistent SSH backdoors. For defenders, agent logs are secrets. Treat them that way.

WIRED reports ransomware claims have spiked, with nearly 3,000 alleged attacks posted to leak sites over the summer, by some counts. Fully AI-powered ransomware hasn't shown up yet. Instead, AI is handling the grunt work. Recorded Future's Allan Liska says groups use it for better phishing emails and for negotiation. It's also lowering the bar to start a crew. ReliaQuest traced one new group to an inexperienced teenager who made more than $800,000 in under two weeks. Does that change our playbook? Not really. Phishing defenses, patching and backups still carry the load.

Token Security lays out a scenario. Let's say someone in accounts builds an AI agent to read an inbox, match invoices in the ERP and post summaries to Slack. Security never knows it exists. Then an invoice arrives with hidden white text asking to change the vendor's bank details. No person reads it. The agent does, and makes the change. In a survey Token commissioned, 82% of respondents had found agents they didn't know about. The rule of thumb is Meta's Rule of Two. Untrusted input, sensitive data, the power to act. Never all three.

OpenAI has launched two security products under its Daybreak platform, according to The Register. Red Team, in early access, uses agents to probe web applications and validate reported vulnerabilities. Defense Factory scans codebases, finds flaws and writes patches, and engineers review and approve each fix before it ships. Greg Brockman, OpenAI's president, said companies will be secure only to the extent of the frontier model they defend with. He also said fundamentals still matter. I'd lean on that second part. Whatever tooling you pick, keep a human on the approve button.

A study from IP protection firm Podqi, shared with The Verge, asked ChatGPT, Gemini and Perplexity where to shop for 151 brands. Across 6,795 answers, all three surfaced risky sites. That means fake storefronts, knockoff sellers, old brand domains someone else re-registered, and dead domains anyone could buy. Perplexity showed them in 15% of answers, ChatGPT in 8% and Gemini in 4%. Smaller brands were hit harder. With the holiday season coming, this is a fraud problem for your users. Teach them to verify the site before typing in a card number.

Australia's government has announced a law that would put mandatory safety rules on frontier AI companies operating there, according to The Decoder. A new oversight body would work alongside the Australian AI Safety Institute, with steep penalties for violations. Model makers would have to allow independent audits, publish their safety frameworks and report security incidents to the government immediately. That last piece is partly about catching early signs of agents going off the rails. For the rest of us, it's a preview. Make sure your incident response plan covers AI systems too.

Three things for this week. First off, treat agent logs and traces as secrets. Scrub credentials out of them and rotate any keys you've shared with AI vendors. Next, inventory your AI agents. Any agent that reads untrusted input, touches sensitive data and can act needs a review. And then lastly, before the holiday rush, remind your users to verify any shopping link a chatbot hands them.

That's the brief. For CompTIA SecAI+ study guides and free security tools, go to skillthropic.com, and subscribe so you catch the next one on Thursday.

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.

Backdoored Claude Code Extension, AI Ransomware Boom and Hijacked Agents | AI Security Brief