Latest / Decrypted: The UK Cyber Briefing / The AI agent flaw a browser tab could exploit, and the Gitea bug already being cryptojacked
Transcript
- 0:03This is Decrypted. Thursday, the twenty-seventh of August.
- 0:07The AI agent flaw a browser tab could exploit, and the Gitea bug already being cryptojacked.
- 0:14Here's a story about what happens when convenience beats a secure default.
- 0:19A vulnerability was disclosed this week in Nvidia's NemoClaw, a wrapper for running local AI coding agents.
- 0:27One visit to a booby-trapped webpage was enough to rewrite the instructions baked into a developer's AI model
- 0:34. Permanently. No phishing, and no credentials stolen.
- 0:39If you're on a UK software team running these agents on developer laptops, to keep code and prompts off thir
- 0:46d-party clouds, pay attention.
- 0:48A sandbox is only as good as the door left open in its wall.
- 0:52Next. What actually broke.
- 0:55NemoClaw runs AI coding agents, of the OpenClaw kind, inside a sandbox called OpenShell.
- 1:03That sandbox is meant to fence off the file system, network and processes the agent can touch.
- 1:10The model runs locally through Ollama, rather than a cloud A P I.
- 1:15To let the sandboxed agent talk to it, NemoClaw started Ollama listening on 0.
- 1:210.0.0. That's every network interface on the machine.
- 1:26It should have been the loopback address, which only the machine itself can reach.
- 1:31And Ollama's A P I had no login and no check on which website was asking.
- 1:36Researchers at Oasis Security, since acquired by Cyera, found that a malicious webpage could exploit that ga
- 1:44p. The trick is D N S rebinding, and it's a well worn one.
- 1:48The attacker's domain first resolves to their own server.
- 1:52Then it flips to 127.0.0.1, and the browser keeps treating the requests as coming from the same site.
- 2:01That was enough to reach the exposed Ollama port and call its model editing A P I.
- 2:07From there, the attacker could rewrite the chat template that sits beneath every conversation the agent has.
- 2:13This isn't the now familiar prompt injection, which affects one exchange at a time.
- 2:19It edits the model's own instructions.
- 2:22The person using it can't see the change, and it survives restarts.
- 2:26And finally. The Secure by Design lesson.
- 2:30The flaw is tracked as C V E twenty twenty-six, six five one zero five.
- 2:36It exists because someone chose 0.
- 2:390.0.0 over loopback to make container to host networking easier, and skipped the authentication that should ha
- 2:48ve gone with that choice.
- 2:54A default that's convenient for the developer, and unsafe for everyone else.
- 2:59Nvidia has patched NemoClaw on macOS and Linux, in version 0.
- 3:050.35. As of 25 August, Windows and W S L users were still exposed.
- 3:12So if your team is experimenting with local AI agents, check now whether your model server is bound to loopb
- 3:18ack only. And don't assume a sandbox around the agent protects the model sitting outside it.
- 3:24That was Decrypted. The written version and every source are linked in the show notes.
- 3:28Thanks for listening.