Behavioural Anomaly Detection
IntentGate baselines each agent's behaviour from its own audit trail, flags deviations, alerts on them, and can revoke the offending token automatically. Enforcement decides each call on its merits; anomaly detection watches the pattern across calls, so a compromise that stays inside per-call policy still shows up as a change in behaviour. A complementary capability alongside the seven-check authorization pipeline.
Why per-call enforcement is not enough on its own
The authorization pipeline evaluates every call and blocks the ones that violate policy. That stops the obviously disallowed action. But a compromised or drifting agent often stays technically in-scope on each individual call while its overall behaviour changes: it starts probing tools it used to leave alone, its call volume jumps, it works at three in the morning, or it accumulates a burst of denials as it tests boundaries. No single call is damning; the pattern is. Behavioural anomaly detection is the layer that reads the pattern.
What IntentGate detects
Detection runs over the gateway's tamper-evident audit trail and computes four signals per agent:
A spike in blocked calls in a short window. Often the first sign of a probing or compromised agent testing what it can reach.
First use of a tool the agent has never called before, relative to its own history.
A sudden jump in call rate against the agent's own established baseline.
Calls outside the agent's normal operating window. Corroborating in combination with the others.
Each signal carries a severity, and every threshold, window sizes, burst counts, spike multiples, the definition of "hours", is tunable per deployment so the baseline fits how your agents actually run.
Detect, alert, respond
Detect
A scan evaluates recent audit activity and produces the signals above, visible in the console so an operator can see which agents are deviating and why.
Alert
Findings are pushed to an external endpoint through a webhook, so anomalies land in the tools your team already watches instead of waiting in a dashboard.
Respond
Optionally, a high-severity signal revokes the offending capability token immediately, cutting the agent off at the gate. Off by default, because revoking a production agent is a deliberate decision.
Built on the audit trail you already have
There is no separate data pipeline to stand up and no model to train. Detection reads IntentGate's own audit chain, which already records every authorization decision with the agent identity, the tool, the decision, the capability token, and the timestamp. The evidence the gateway produces for compliance is the same evidence that feeds detection, which keeps the capability lightweight and keeps its findings explainable, every flag points back to specific, verifiable audit rows.
What anomaly detection does and does not do
It surfaces behavioural change and can act on it; it is not a general-purpose SIEM or UEBA replacement. For enterprises that already run a SIEM, IntentGate's audit trail exports cleanly (Splunk, Datadog, Sentinel, S3, webhook) so the same signals can be correlated with the rest of the estate. Anomaly detection gives you a fast, agent-aware first line inside the gateway; the SIEM remains the system of record for cross-domain correlation.
Related capabilities
Anomaly detection consumes the authorization pipeline's audit output and acts through capability-token revocation. See the Standards Alignment page for how continuous monitoring and response map to common compliance frameworks.
Frequently asked questions
What signals does IntentGate detect?
Four signals over the audit trail: denial bursts, first use of a new tool, volume spikes against the agent's baseline, and off-hours activity. Each has a severity, and every threshold is tunable per deployment.
Can it respond automatically, or only alert?
Both. By default it detects and alerts through a webhook so a human decides. Optionally, automatic response revokes the offending capability token on a high-severity signal, cutting the agent off immediately.
Does it need a separate data pipeline?
No. Detection runs over IntentGate's existing tamper-evident audit trail, so there is no separate log shipping or model training step to operate.
See it in the live demo → Explore the audit chain and anomaly scenarios on the demo page.