NeuralTrust | The leading security platform for generative AI

Behavioral-security detection looks beyond a single request to the actor’s behavior over time. The anomaly_detector detector scores each actor (keyed on consumer_id) for bot-like timing, repeated payloads, repeated failures, and abuse spread across collectors.

Property	Value
Slug	`anomaly_detector`
Category	behavioral_security
Sides	input
Protocols	all
Mutable	— (detection‑only; enforcement set on the policy rule)
State	Stateful — tracks history per `consumer_id` (Redis)

Because it tracks per-actor state, always pass a stable consumer_id on the guard request (e.g. the authenticated user or API client). Without it, behavioral scoring can’t attribute activity.

How it scores

The detector combines three weighted signals into a single score in [0,1]:

Timing pattern — unnaturally regular or rapid request cadence (bot-like).
Content similarity — repeated or near-identical payloads (scripted abuse).
Cross-collector — the same actor hitting many collectors/keys.

The weights must sum to 1.0. A finding is reported above threshold.

Settings

Field	Type	Default	Notes
`threshold`	number	`0.7`	Score above which the actor is flagged.
`retention_period`	integer	`300`	Seconds of actor history retained.
`min_time_between_requests`	integer	`1`	Seconds; faster cadence raises the timing signal.
`max_requests_to_analyze`	integer	`10`	Sample window size.
`max_failures`	integer	`5`	Failure count that contributes to the score.
`cross_collector_max`	integer	`5`	Distinct collectors before cross-collector signal saturates.
`timing_pattern_weight`	number	`0.33`	Must sum to `1.0`…
`content_similarity_weight`	number	`0.33`	…with the other two weights.
`cross_collector_weight`	number	`0.33`

{
  "name": "Flag abusive actors",
  "plugin_slug": "anomaly_detector",
  "settings": {
    "threshold": 0.7,
    "retention_period": 300,
    "min_time_between_requests": 1,
    "timing_pattern_weight": 0.4,
    "content_similarity_weight": 0.4,
    "cross_collector_weight": 0.2
  }
}

The detector reports a finding (signal.type: "anomaly"); the policy rule decides whether that Monitors or Blocks.

When to use

Pair with rate limiting at TrustGate — rate limits cap volume, anomaly_detector catches patterns that stay under the cap.
Start with the Monitor action and tune the weights/threshold against real traffic before switching the rule to Block, since behavioral signals are probabilistic.
Requires a stable consumer_id across requests to be effective.

​How it scores

​Settings

​When to use

How it scores

Settings

When to use