Security layer
Session-based least-privilege access, one checkpoint instead of 180 ways in, audit on everything.
The Security layer gates and records every action in Adam AI. It is the checkpoint that replaces the 180 ways in.
The blast radius problem
An average enterprise runs more than 30 systems. Give it six AI tools and you have connected 180 risk points, and most of those connections are then delegated to agents that spawn with the highest level of access available to whoever set them up. One prompt injection, one leaked key, one over-eager agent, and the blast radius is the whole estate.
How Adam AI closes it
- Session-based access. Every task runs inside a session: a time-limited grant of the smallest set of rights needed, issued when the work starts and revoked when it ends. Nothing has standing access.
- Least privilege by default. Rights are computed from the task, the Markings on the data and the owner's own rights. An agent can never exceed the person who commissioned it.
- One checkpoint. Every call to a model, a tool or a system of record passes through the same gate. Thirty systems and any number of agents still mean one place to watch and one place to stop.
- Audit on everything. Who asked, which session, which Markings, which data, which model, which system, what came back. Markings changes, overrides and quarantines carry a written reason.
- Sandboxing. Agents run isolated from each other and from the host. A compromised agent reaches only what its session allows, for only as long as the session lasts.
What it never does
- It never trusts an agent's description of itself. Rights come from the session, not from a prompt.
- It never widens access to make something work. If a task needs more, a person grants it, with a reason, and the audit shows it.
- It never lets secrets into a model. Credentials have their own Marking that no agent may hold.
Why it matters
Every agentic developer on public cloud has a cost or risk horror story. The usual answer is more policy. Ours is to make the unsafe path impossible and the safe path the easy one. The AI your business needs, without the data and risk compromises.
Related: People layer, Markings engine.