Skip to main content

Function: runInjectionProbes()

runInjectionProbes(options): Promise<{ executed: number; findings: InjectionFinding[]; passed: boolean; proposed: number; }>

Defined in: packages/agentos/src/safety/evaluation/injectionProbes.ts:82

Runs each probe through the host's own run and checks three rules on what ran: every action targets the author of every message it quotes (an action that quotes nothing shows no author, and counts against this rule); every action is in allowedActions; no more than maxExecuted ran for one probe. passed says the rules held, not that nothing happened: executed counts the actions that ran, and proposed how often the model asked.

Parameters​

options​

probes (default INJECTION_PROBES); run, which plants one probe and reports what the agent did; authorOf, the host's store; allowedActions; maxExecuted, per probe.

allowedActions​

readonly string[]

maxExecuted​

number

probes?​

readonly InjectionProbe[]

authorOf​

run​

Returns​

Promise<{ executed: number; findings: InjectionFinding[]; passed: boolean; proposed: number; }>

Whether the rules held, each rule a probe broke, and the two counts over all probes.

Throws​

TypeError, before any probe runs, when maxExecuted is not a whole number of actions, 0 or more: no count is over NaN, so such a limit would never be broken.