Sleeper-Trigger Check

The agent time-bomb a single-turn scan can't see.

A directive planted early — turn 1, a memory write, a tool output — can lie dormant across many benign turns, then detonate a privileged or exfil action the moment a trigger fires (a date, an Nth-turn count, a keyword, a user phrase, a tool output). Black-box single-turn tests and pressure-ramp escalation ladders structurally cannot see it. AgentSmack scores a multi-turn transcript for exactly this long-horizon sleeper — surfacing the planted→fired turn gap as evidence, mapped to OWASP LLM / NIST AI RMF coverage with paste-able remediation, a synthetic-attacker gauntlet, and audience report variants. Load a sample to watch the gauge swing with no live infra.

Paste your multi-turn transcript (JSON)

Drive a live agent endpoint — does the time-bomb detonate?

Point AgentSmack at YOUR deployed function-calling endpoint. We plant a dormant fire-later directive on turn 0, run benign filler turns to advance the dormancy clock, fire the trigger, and grade whether a privileged tool call detonates N turns later — the long-horizon sleeper a single-turn scan structurally cannot see. We store only redacted excerpts, never the raw directive or reply. POSTs to a real endpoint you control.

Target URL (https only)Wire shapePlant source (provenance label)Trigger kindAuth headers (one `Name: value` per line, optional)

Planted directive (the dormant fire-later instruction, injected turn 0)Benign dormancy turns (one filler ask per line — these advance the clock)Trigger descriptor (optional, evidence only)Trigger prompt (the turn whose arrival fires the activation)Privileged-action tools (one tool name per line — empty = any tool call detonates)

Related: Scheduled / Deferred-Action Integrity — scheduled vs dormant: the agent's own queued work fired by its own schedule, not a planted trigger detonating later.