Tools · Security · 12

Trace what a prompt injection can reach before assigning a risk score.

The problem does not end when a model interprets an untrusted instruction. Impact depends on which data, tools, outbound channels and memory the influenced model can reach next. This explorer models those paths and the controls that cut them, without generating attack payloads or teaching offensive techniques.

Inputdirect, indirect or multimodal
Boundaryisolated reader vs privileged model
Impactdata, actions, egress and memory
Methoddeterministic reachability, not probability

Scenario

Surface and privileges

Independent controls

Modeled posturenot a probability
Reachable pathsout of 5
High-impact pathsconfidentiality, integrity or persistence
Privileged influence

Impact paths

Instruction steering

Known-secret leakage in rendered output

Consequential action without independent approval

Data exfiltration through external egress

Persistent memory poisoning

Controls in this scenario

Isolated readerisolation

For indirect content, the component that reads untrusted data should not hold the same authority as the privileged actor.

Least privilegeimpact

Reduces available damage after steering; it does not block a binary path by itself or prevent injection.

Intent validationaction

Checks proposed actions against the original task outside the untrusted-content path.

Human confirmationaction

Places an independent approval boundary before sending, deleting, purchasing, publishing or modifying sensitive state.

Egress restrictionegress

Limits destinations and channels capable of moving information beyond the trust boundary.

Memory validationpersistence

Stops influenced content from automatically becoming reusable state in future turns.

Secret filteroutput

Can block known secret classes; it does not make every sensitive datum detectable and is only one defensive layer.

Limit: this tool evaluates architectural reachability assuming enabled controls are enforced correctly. It does not estimate the probability that a model follows malicious instructions or the real-world failure rate of any control.

Sources and method

Impact mitigation needs boundaries outside the prompt.

OWASP distinguishes direct and indirect injection and recommends content separation, least privilege, output validation, agent-specific controls and human approval for privileged operations. OpenAI likewise recommends limiting access and reviewing consequential actions.