Submission + - Autonomous Agents Need Mandatory Access Control (Agents gone wild...) (legiscope.com)
exodude writes: Everyone has seen their own version of “Agents Gone Wild” by now: from deleted databases to leaked credentials, chatbots always justify their action saying they were trying to help. While you’re still hallucinating about the results—and staring at the $20,000 credit card bill your agent just ran up—we decided to write a technical proposal to put agents on a (technical) leash so we don't have to find about “what’s next.”
The proposal draws on FreeBSD’s MAC framework, Capsicum and firewalls. Agent API requests would carry mandatory security policies defining which IP addresses agents can contact, which files they can access, which processes and tools they can use, and what permissions their children inherit. The runtime and tool services enforce these rules outside the model and reject requirements they cannot enforce. OpenAI and Claude already provide useful building blocks; the proposal brings them into a common contract covering the entire run.
The proposal draws on FreeBSD’s MAC framework, Capsicum and firewalls. Agent API requests would carry mandatory security policies defining which IP addresses agents can contact, which files they can access, which processes and tools they can use, and what permissions their children inherit. The runtime and tool services enforce these rules outside the model and reject requirements they cannot enforce. OpenAI and Claude already provide useful building blocks; the proposal brings them into a common contract covering the entire run.