v0.4.0GitHub

Definitions · A

Agentic Privilege Escalation

Agentic privilege escalation occurs when an AI agent obtains or uses capabilities beyond its intended scope — often via prompt injection, tool chaining, or leaked credentials in context. Runtime policy and lineage controls reduce blast radius.

Agentic privilege escalation happens when an AI agent gains or uses capabilities beyond its intended scope — for example executing destructive SQL after reading an injected document, or calling admin tools from a read-only workflow.

Mitigations

  • Runtime tool policy (allow/deny/mask)
  • Lineage contamination after untrusted content
  • Trust Vault tokenization so secrets never enter model context

See Lineage Firewall and the threat model.

Related terms

← All definitions · Runtime simulator