RESPONSIBLE AI

Governance where the agent acts, not where it thinks.

Most Responsible AI work is about the model: fairness, evaluation, training data, transparency of outputs. Aegis sits at a different, concrete point: the moment an AI agent tries to call a tool or read data through MCP. At that boundary it enforces human oversight, least privilege, auditability and data minimization. That makes it a practical control for the accountability and safety parts of a Responsible AI program.

THE BOUNDARY

Aegis owns the action and data-access slice.

A Responsible AI program spans the model, the data, and the actions the system takes. Aegis does not touch the first two. It owns the third: every tool call, resource read, prompt and completion that an agent routes through MCP passes a fail-closed decision plane that applies identity, policy, approval, egress, screening, output control and audit before anything reaches a real system.

Principle served: an AI system should not be able to take a consequential action, or expose sensitive data, without an enforced, recorded, reviewable control. Aegis is the enforcement point for that principle on the tool path.

HUMAN OVERSIGHT

A person stays in the loop for consequential actions.

Independent approval

Destructive and mutating tool calls require a per-call approval from a different identity than the requester. The agent cannot approve its own action, and a delegated identity cannot stand in for the approver. This is meaningful human oversight at the point of action, not a policy document.

Dry-run receipts

Before a destructive call commits, Aegis produces a receipt of what will happen, bound to the caller, tool, arguments and definition digest. The approver sees the exact action they are authorizing, and the binding prevents a swap between approval and execution.

ACCOUNTABILITY

Every decision is traceable and contestable.

PRIVACY

Data minimization at the tool boundary.

Secret and PII redaction

Results are scanned for credentials, keys, PEM blocks and personal data patterns, which are redacted before the model sees them. High-entropy strings with no recognizable prefix can be redacted too.

Handle-only output

For tools whose output cannot be safely released, Aegis returns an opaque handle rather than the content, so raw or uncertain data never enters model context.

Least privilege

Tool classes, typed resource scopes and per-caller token exchange give each call only the access it needs, so an over-broad agent cannot reach a tenant or a dataset it was not scoped to.

No token passthrough

A caller's bearer token is never forwarded to a backend. The backend receives only a scoped, gateway-issued credential, which limits how far a compromised agent can reach.

NIST AI RMF

Mapping to the four functions.

FunctionAegis contribution
GovernPolicy as configuration, separation of duties through approver groups, fail-closed defaults, and a documented threat model and security properties.
MapA typed catalog of exactly which tools, resources and destinations an agent may reach, with each tool classed by its risk surface.
MeasureA hash-chained audit journal, decision records with reasons, and a repeatable control suite plus the before-and-after CVE containment run.
ManageIndependent approval and dry-run receipts for consequential actions, live token and session revocation, rate limits and budgets, and break-glass containment of a source.

The NIST AI Risk Management Framework names these four functions. Aegis addresses the action-and-data parts of each; model-centric activities such as bias measurement stay with the model team.

EU AI ACT

Concrete touchpoints for high-risk obligations.

Human oversight

The independent-approval and dry-run-receipt controls are a direct, enforced mechanism for keeping a person able to intervene before a consequential automated action, which the Act requires for high-risk systems.

Logging and record-keeping

The tamper-evident audit journal, with intent recorded before execution, supports the Act's record-keeping expectation: an operator can reconstruct what the system did, when and under whose authority.

These are enabling controls, not a compliance certification. The EU AI Act obligations fall on the provider and deployer of the overall AI system; Aegis gives them an enforcement and evidence layer for the action boundary. Confirm applicability with your own compliance function.

HONEST LIMITS

What Aegis does not do.

REFERENCE MAP

Continue in the guides.