Least access, always
Every agent starts with nothing and gets only what its role’s policy grants.
Our security approach
Assume the agent will try. Make sure it can’t.
Agents can be wrong, manipulated or misused. RayShield is built so safety doesn’t depend on the agent behaving well.
Least accessDeterministicTamper-evident
Four principles
Enforce where the agent can’t reach
Controls sit below the agent’s environment, out of its sight and out of its control.
Deterministic over probabilistic
Allow or block, every time, by policy. No silent fallbacks.
Evidence for everything
Tamper-evident records of what ran, what it touched and what was stopped.
What Aegis enforces
Isolated executionDefault-deny networkingRuntime-only secretsApproved-only connectorsData loss preventionControlled policies and logs
What we don’t claim
Prompt injection
We don’t claim to stop prompt injection outright. We shrink what a manipulated agent can reach and record what it tried.
Coverage follows installation
Agents running outside Aegis are detected at the network level, not contained. We show you exactly where that line is.
A compromised machine
Aegis narrows the damage and protects the evidence, but it doesn’t make a hijacked device trustworthy.