Definition
Behavioral Baseline Integrity Score (BBIS) measures how well behavioral baselines resist simulated slow-and-low adversarial adaptation.
The Measurement Gap
A financial services firm's security team uses an ITDR platform for behavioral anomaly detection. The platform learns each user's normal behavior (sign-in times, location patterns, application access patterns, data access patterns) and flags deviations. Detection rates on known threat patterns are excellent. The security operations team holds the platform up as evidence of mature threat detection.
The firm engages a red team for a periodic adversarial assessment. The red team simulates an advanced persistent threat scenario: compromise a service account through credential theft, then patiently expand access through small increments that look like normal organizational evolution. Over the course of the simulated 60-day campaign, the red team makes incremental privilege grants, accesses adjacent systems gradually, and shifts the compromised account's behavior pattern day by day in ways that each individual day would be within the baseline's tolerance. By day 45, the compromised account has accessed sensitive data repositories that would have generated immediate alerts if accessed on day 1. By day 60, the behavioral baseline has shifted to treat the compromised behavior as normal.
The ITDR platform generated no alerts during the campaign. The detection that worked on known threat patterns failed on the patient adversarial adaptation pattern. The baseline learned that the compromise was normal because the changes were small enough each day to be absorbed.
The legacy identity metrics report a healthy detection program.
Authentication success rate: stable.
MFA enrollment: 100%.
Anomalous access alerts: stable rate.
Mean time to detect anomalous access: low, in line with industry benchmarks.
False positive rate: low.
Threat coverage: tracked qualitatively, considered comprehensive.
None of these report on whether the behavioral baseline itself can be subverted by patient adversaries. The detection metrics measure what the system detects. The patient-adversary failure mode is what the system does not detect, and what the system does not detect is not in the detection metrics.
These metrics were designed when threat detection used predefined rules and signature-based patterns. The patterns were updated by humans who had explicit threat intelligence. With AI behavioral baselines that learn from observed activity, the baselines themselves become an attack surface: the adversary who can teach the baseline what is normal can operate under it indefinitely.
There is no metric that tells you whether an attacker has contaminated the behavioral baseline used for anomaly detection, which means detection rates on known patterns look strong while baseline-adapted adversaries operate undetected.
Formula
Behavioral Baseline Integrity Score (BBIS) measures how well behavioral baselines resist simulated slow-and-low adversarial adaptation. In plain language: run a red-team exercise that simulates patient adversarial escalation across multiple stages. For each stage, observe whether the behavioral analytics generates an alert. BBIS is the percentage of escalation stages that triggered detection. A high value means the baseline detects patient adversaries. A low value means the baseline absorbs them as normal.
The formula:
BBIS = Escalation_stages_detected / Escalation_stages_total × 100
Where:
- Escalation_stages_detected = simulated escalation stages where the behavioral analytics platform generated an alert
- Escalation_stages_total = total stages in the simulation (recommend 10 stages spanning weeks)
Units are a percentage from 0 to 100. 100% means every escalation stage was detected. 0% means the baseline absorbed the entire adversarial progression. A worked example. The financial services firm runs BBIS via a red-team exercise spanning 60 days with 10 distinct escalation stages on a simulated compromised account:| Stage | Day | Activity | Alert? ||-------|-----|----------|--------|| 1 | 3 | First abnormal off-hours login | YES || 2 | 8 | Access to adjacent department's shared drive | NO || 3 | 14 | Privilege grant request via legitimate workflow | NO || 4 | 19 | Access to first sensitive repository | NO || 5 | 25 | Increased volume of database queries | NO || 6 | 31 | Access to financial reporting tool | NO || 7 | 37 | Data export from financial reporting | NO || 8 | 43 | Access to executive communication channels | NO || 9 | 51 | Email rule creation forwarding to external | YES || 10 | 58 | Bulk credential file access | NO |2 of 10 stages detected.
BBIS = 2 / 10 = 20%
Per the catalog guidance (above 70% resists most adaptation, below 40% patient adversary could operate undetected) this BBIS of 20% is severely below threshold. The action trigger is to implement baseline stability controls: baseline anchor points that don't drift, deviation alerts based on absolute behavior rather than baseline-relative behavior, and explicit detection rules for the escalation patterns the simulation revealed. Three edge cases need explicit treatment. First, simulate realistic adversarial patience: days to weeks between escalation steps. A simulation that escalates over hours will trigger easily and overstate BBIS. Second, different user profile types have different baseline stability: service accounts vs. human users, executives vs. individual contributors, frequently-changing roles vs. stable roles. Run simulations against representative profile types and report BBIS per profile. Third, the minimum is 5 red-team exercises per year, ideally with rotating attack patterns. Cross-references. Upstream of BBIS is
Telemetry-to-Mitigation Coverage Drift (TMCD, GMI-IT-011): behavioral baseline drift is a specific case of telemetry-to-mitigation drift. Downstream is
Adversarial Threat Coverage Index (ATCI, GMI-IT-010): baseline integrity directly determines the system's ability to detect sophisticated adversaries. The misinterpretation to watch for: a team reports BBIS = 80% (above the threshold) and concludes the baseline is robust. They miss that the simulation used escalation patterns the security team had defended against in previous exercises, so the patterns were familiar to the baseline (effectively in the training data). Rotate simulation patterns across exercises and treat BBIS as meaningful only against patterns the platform has not been tuned against.
Proposed formula, pending validation pilots.