Threat intelligence

Learn from every agent, defend each one.

Reputation the whole network contributes to, model-scored threat classes your rules can target, and a behavioral baseline per agent. Each layer is fail-open, so intelligence sharpens decisions without ever blocking on them.

Cross-tenant feed · model-scored classes · per-agent baselines

network judgment
host tools.shadyvendor.io
flags 12 organizations
risk 3590
fail-open: intelligence raises risk, never blocks on its own
0
Intelligence layers
~20
Calls to learn a baseline
Spike threshold
k ≥ 5
Anonymity floor

How it fits together

Three layers, one decision.

Each layer feeds the same risk picture. Together they catch what a single call never could.

Cross-tenant reputation

An opt-in feed of card verdicts, protected by a k-anonymity floor and fail-open by design.

Model assessment

A model scores your riskiest calls and returns a threat class your rules can target.

Behavioral baseline

Per-agent normal for tools, risk, and active hours, with deviations flagged.

Network reputation

When one tenant flags a host, everyone benefits.

Verdicts are recorded anonymously with a k-anonymity floor. Once enough organizations have flagged a host, that shared judgment folds into your local risk on later preflights.

k-anonymity floor at 5

Most reports were a deny or high score
Feed unavailable

Network judgment

tools.shadyvendor.io

your preflight risk3590

Network raises your risk to 90

12 organizations have flagged tools.shadyvendor.io. That shared judgment folds into your local risk as critical, so your agent treats the host with the caution the wider network already learned.

Model-scored

A second opinion on your riskiest calls.

High-risk and critical calls are assessed by a model that returns a threat class and a calibrated score your policies and response rules can target.

Team and above
A risky call
aws.s3.deleteBucket96
model assessment
threat class
data exfiltration
calibrated score
96
recommendation
block

Use it in a rule

Policies and response rules can target data_exfiltration directly. Policy engine and automated response both accept a threat class as a predicate.

Behavioral

Every agent has a normal. We learn it.

After about twenty calls the platform knows an agent's usual tools, risk, and active hours, then flags a new tool, a risk spike beyond mean plus three sigma, and off-hours activity.

Observed call

baseline: 84 calls

normal band 0 to 52 (mean + 3σ)

usual hours 13:00 to 21:00 UTC

Still learning this agent

What the baseline flags

  • medium
    new tool

    first-ever call to "aws.s3.delete"

  • high
    risk spike

    risk 80 far above its baseline (mean 22)

  • low
    off hours

    active at 03:00 UTC, outside its usual hours

Turn every agent's behavior into a defense.

Share what the network learns, score your riskiest calls, and catch the agent that suddenly acts unlike itself.