Guides · ← All guides
AI Agent Velocity Anomalies: When Speed Itself Is the Attack Signal
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 7 min read
The short answer
Agents have rhythms, and attacks break rhythm before they break rules: a loop multiplies call rate, a sweep multiplies record rate, an injection multiplies both at hours the agent never works. Velocity detection baselines each agent's rates — calls, records, spend, failures — and treats deviation as a first-class event: Rogue-Agent Defense ships velocity, burst-rate, off-hours-spike and fail-streak guards because rate is the one signal that fires during the expensive minute, not after it. The responses scale with confidence: a soft deviation holds new actions for review, a hard one quarantines — the ladder in quarantine practice driven by arithmetic instead of vibes.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works while you read — most controls in this guide start working the moment a canary, a policy check, or a spend cap is in place.
What is a velocity anomaly in an AI agent? a deviation from an agent's established rate pattern — calls per minute, records per hour, spend per run — large or sudden enough to signal a loop, a bulk collection, or an injection steering the agent at machine speed.
Baseline the rhythm, per agent
A fleet-wide threshold is useless: the research agent's busy hour is the bookkeeping agent's incident. Baseline per agent over its first weeks — calls per hour by time of day, records per run, spend per run, failure rate — and keep the baseline as a living distribution, refreshed monthly, because jobs legitimately grow. The point is not a magic number; it is a defensible statement of normal that a deviation can be measured against.
Seasonality is the trap that kills naive baselines: month-end agents, campaign agents, Monday-morning agents. Their baselines need the calendar baked in — rate expected for this kind of day — or the guard pages on schedule and gets muted, which is worse than no guard, because a muted channel stays muted on the real day too.
The four rate guards and what each one catches
Velocity: sustained rate above baseline — the loop and the patient sweep. Burst-rate: a short window at extreme rate — the injection that wants everything now, before anyone wakes. Off-hours spike: activity when the agent's job sleeps — injections do not respect rosters, and legitimate work rarely starts at 03:00 unannounced. Fail-streak: consecutive failures with retries — not an attack signal but a cost and signal-integrity one, from the spend story, because a failing agent at full rate is burning money to produce errors.
Each guard deserves its own threshold and response. Burst is a hold-now signal; fail-streak is a circuit breaker; off-hours is a context multiplier that lowers every other guard's threshold until morning. Composing guards this way is what keeps the page count low and the pages meaningful.
Respond in a ladder, not a binary
Binary alerting — page or nothing — trains teams to choose nothing. Build the ladder: deviation ×2 logs and tags the run for morning review; ×5 holds new actions in APPROVE and notifies the owner in working channels; ×10 or any burst auto-quarantines and pages. The exact multipliers matter less than the property: response proportional to deviation, with the top rung fully automatic, because the top rung is the one that fires at machine speed against a machine-speed problem.
The ladder also protects legitimate surges: the marketing agent on launch day climbs one rung, gets reviewed, and its baseline learns the launch pattern with a human annotation. Annotate, do not just clear — the annotation is how the system distinguishes the second launch (expected, logged, no page) from the first (surprising, correctly).
Rate as the tripwire for the quiet attacks
Velocity's deepest value is against attacks designed to be quiet. The slow exfiltration that respects per-action thresholds still bends the cumulative curve — records per week drifting up across a month — which a rolling-window guard sees even when every individual day looks innocent. This is the same cumulative logic as the volume budgets in bulk export defence, generalised: any resource an attacker needs — calls, records, spend, approvals — can be metered on long windows.
Pair rate with identity: the same totals mean different things from the agent's usual host, session pattern and tool mix. A rate spike composed of a new tool is the new-tool-baseline story and should be held regardless of rate; a rate spike of the usual tool at the usual hour is capacity planning. Guards that read composition, not just counts, page people about attacks instead of about success.
Tuning without silencing
Every guard produces a tuning log: fires, dispositions, threshold changes. Review it monthly with one rule — a threshold may be raised only with a named reason (job grew, launch calendar, new market) recorded beside it. Thresholds raised silently during a noisy week are how a detection system dies with its configuration intact; the postmortem discipline applies to guard changes as much as to incidents.
And measure the guard system itself: pages per month, fraction actioned, catches made. A healthy system pages rarely and is believed completely. Belief is the real asset — the night the burst guard fires on a genuine sweep, the on-call engineer must act first and doubt later, and that trust is built by every previous page having been worth reading.
Five actions to take today
- Baseline per agent. Calls, records, spend and failures by hour and day type, refreshed monthly from clean runs.
- Deploy the four guards. Velocity, burst-rate, off-hours spike and fail-streak, each with its own threshold and response.
- Build the response ladder. Log at ×2, hold and notify at ×5, auto-quarantine and page at ×10 or any burst.
- Add long windows. Rolling 30-day rates per agent, so slow campaigns bend a curve someone is watching.
- Log every threshold change. A named reason beside each adjustment, reviewed monthly — no silent raises.
A worked example: eleven minutes at Vantage Property Group
The following is a fictional example built from real threat classes and controls. The company and numbers are invented; the mechanics are not.
Vantage Property Group, a 38-person agency, runs a listings agent that syncs property data and drafts descriptions: baseline 40–70 calls a day, working hours only, records per run under 100. At 02:51 on a Sunday the burst guard fires: 300 calls in 4 minutes, a rate 25 times the hourly baseline, composed of read calls against the full listings table at maximum page size.
The ladder climbs itself: burst means auto-quarantine and a page, no human decision required at 02:51. The frozen trail shows the cause by breakfast — a compromised partner feed whose records carried an instruction block telling the agent to inventory all listings and prepare them for migration. Records actually read before the freeze: 1,900 of 11,400; records sent anywhere: zero, because the reads were the attack's first phase and the send phase never got its turn.
The tuning review finds one improvement and resists one temptation. Improvement: partner-feed content now routes through a quarantined reader. Temptation, resisted: several legitimate daytime bursts had brushed the ×5 rung that month, and the lazy fix was raising thresholds; instead the launch calendar was annotated into the baseline, keeping the guard sharp for the next 02:51.
Frequently asked questions
What is a velocity anomaly in an AI agent?
A deviation from the agent's baselined rate pattern — calls per minute, records per hour, spend per run, failure rate — that signals a loop, a bulk collection or an injection at work. Rate fires during the expensive minute, which makes it the earliest reliable signal for several attack classes.
Which rate guards should an agent have?
Four, each catching a different shape: velocity for sustained above-baseline rate (loops and patient sweeps), burst-rate for short extreme windows (grab-it-now injections), off-hours spike for activity when the job sleeps, and fail-streak for expensive retry behaviour. Each needs its own threshold and response.
How do you set velocity thresholds without false pages?
Baseline per agent with the calendar baked in — expected rate for this kind of day — and respond in a ladder: log at small deviations, hold and notify at medium, auto-quarantine only at large deviations or bursts. Annotate legitimate surges into the baseline rather than raising thresholds, and review every threshold change with a named reason.
Can velocity detection catch slow, quiet exfiltration?
Yes, on long windows: a drip campaign that respects every per-action threshold still bends the cumulative curve — records per week drifting upward across a month. Rolling 30-day metering per agent, per resource, is the control; individual days can all look innocent while the month cannot.
What should happen when a burst is detected?
Auto-quarantine and page, without waiting for a human decision — bursts are machine-speed events and the top rung of the ladder must be automatic. The frozen trail then scopes what the burst achieved; in the worked example, a 25-times-baseline burst at 02:51 was frozen with zero records sent.
Related guides
- How to Stop an AI Agent Bulk Export — the campaign budgets behind the rate story.
- AI Agent Spend Caps — rate applied to money: spikes and fail-streaks.
- How to Stop an AI Agent Looping Forever — the velocity signature's most common cause.
- How to Quarantine an AI Agent — the top rung of the response ladder.
- What Is a Rogue AI Agent? — velocity among the canonical rogue signs.
- AI Agent Audit Trails: What to Log — the baselines' raw material.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works, and let the next rogue action announce itself.