AI WORLD HQ
Imagine an AI assistant that can read your work email, check a shared calendar, and place an order for supplies. That kind of helper is useful, but it also raises a simple question: who decides what it is allowed to do? IBM's latest research post focuses on that question.
What the announcement covers
According to IBM, its research team uses AI across its own operations and wants the benefits without giving up safety, security, or cost control. To support that, IBM says it works on shared tools that teams can test and improve together.
The main piece is Praxis, an open-source proxy server. A proxy is a middleman that sits between a user's request and the service that answers it, so it can inspect and manage traffic. IBM worked with Red Hat to contribute the Praxis Policy Engine, which Praxis uses as a filter for both ordinary web traffic and AI traffic. IBM also says Praxis has been accepted into the Cloud Native Computing Foundation (CNCF) Sandbox, an early-stage project program under that open-source foundation.
How the policies work
IBM describes PPE as treating several things as "policy targets," meaning things that rules can be written for:
- Tools: actions an AI can take, such as searching a database
- Prompts: the instructions people type to an AI
- Model calls: requests sent to the AI model itself
- Other agents: one AI system asking another for help
Teams can write hard access rules ("this agent may never see payroll files"), softer behavior guidelines, and rules for specific fields, all in one policy file. IBM says this can run at a gateway, beside an individual agent, or inside a development framework. Each placement covers the actions that pass through it.
IBM also says PPE can write an audit record for each decision it makes. An audit record is a log entry that shows what was allowed or blocked, and when. Optional signatures and "hash chaining," a method that links each log entry to the one before it, are meant to help someone spot if records were changed later.
Why it matters
The post explains that AI gateways are becoming their own layer in AI systems. Think of a gateway as a security desk at a building entrance: it checks who is coming in, what they are asking for, and keeps a record. IBM says gateways also handle access control, metering of how much AI text each user or team consumes, and parts of how agents carry out multi-step tasks.
IBM says it is working on related standards, including an agent identity standard with CoSAI, and is studying how controls on a computer and controls on a network could work together. Those are research goals, not finished products.
Wrap-up
IBM's announcement is about building the referee before the players get too powerful: rules, identity checks, and records that run alongside AI agents. For now, it's an open-source research effort that developers and security teams may follow. A useful next step today is simple: open the AI tools you already use and check which accounts and files they are allowed to reach.