Build on the supported package contract
Use @absolutejs/agent-trust through its supported public entry points.
@absolutejs/agent-trustv0.2.2betaAIProvider-neutral provenance, taint propagation, instruction/data separation, and sink policy enforcement for AI agents.
bun add @absolutejs/agent-trustProvider-neutral provenance and taint tracking for agents. Every value carries its purpose, authority, source, digest/proof, taints, parent digests, and evidence-bearing transformations.
The central security rule is structural: external content is data, never an instruction merely because it contains imperative language. compileAgentContext returns separate instruction and data segments instead of interpolating tool, web, file, or memory content into a privileged prompt. Sink policies then fail closed before instructions or actions cross a trust boundary.
Use compileGuardedAgentContext at provider boundaries. It enforces the instruction policy before compilation; the basic compiler also conservatively downgrades externally authoritative or tainted values mislabeled as instructions into the data channel.
Taints propagate through derived model output. A sanitizer may remove named taints only while recording evidence and a new digest; there is no implicit "trusted because the model said so" conversion. Proof verification is injected, so signed discovery descriptors, DSSE, JWTs, or provider KMS can share the same contract.
Outcomes
Use @absolutejs/agent-trust through its supported public entry points.
Hardening checklist
Follow in order
Supported entry points declared by this package manifest.
Package entry point declared in package.json.
Scripts declared by this package manifest.
Search the declarations exported by the current package type files. Expand a symbol to inspect its source-backed signature.