Teams set rules for their AI but never learn how much pressure breaks them.
Attacks each rule with step-by-step adversarial pressure and reports exactly where it fails.
An installable package (pip install bct-framework) that any project can run.
The problem
An AI system may be assigned an explicit behavioral constraint — never disclose unmasked financial data, never provide a direct answer — yet organizations typically lack a systematic method for determining the precise threshold of pressure required to breach that constraint, or which specific attack pattern breaches it first.
Who it is for
- Organizations building AI systems with mandatory behavioral constraints — compliance boundaries, pedagogical requirements, safety limits — requiring empirical knowledge of actual failure conditions, not assumed resilience
- Security and QA organizations requiring evaluation of system-specific defined constraints, beyond generic vulnerability classifications
Market context
This is an active category shaped by material regulatory pressure — EU AI Act adversarial testing requirements for high-risk systems entered into force in August 2026. Established tooling includes Garak (NVIDIA's probe-based scanning infrastructure), PyRIT (Microsoft's multi-turn attack orchestration), Promptfoo (OpenAI-owned, CI-native), and DeepTeam (OWASP-aligned red-teaming framework).
The gap it closes
Established platforms predominantly test against known, catalogued attack taxonomies — jailbreak patterns, prompt injection, standardized probe libraries. Few are architected specifically to evaluate a system's own custom-defined behavioral constraints under incrementally escalating pressure, quantifying the precise threshold and methodology at which a specific constraint fails.
How it works
- The specific behavioral constraint requiring validation is defined
- BCT applies systematically escalating adversarial pressure, targeting that constraint specifically
- The precise failure point is reported — pressure level and methodology that breached the constraint
- Deployed as a standalone, installable package, enabling identical tested logic across multiple projects
What makes it different
- Tests system-specific behavioral constraints, not generic, pre-catalogued attack signatures
- Quantifies the precise failure threshold — reporting not merely pass or fail, but the exact pressure level and methodology at which a constraint yields
- Architected as reusable infrastructure — extracted into a standalone package specifically because duplicated adversarial-testing logic constitutes an independent quality liability
Architecture
System Overview
Component Breakdown
| Rule Definition Interface | Specifies the exact behavioral constraint under test |
| Graduated Pressure Engine | Applies escalating adversarial pressure in defined stages |
| Breaking-Point Reporter | Identifies the exact stage and method that first caused failure |
| Standalone Package | pip install bct-framework — reused identically across projects |
Data Flow
- A specific behavioral rule is defined for the target AI system
- BCT applies pressure starting mild, escalating in defined stages
- At each stage, the response is checked against the rule
- The first breaking stage and method is recorded and reported
Technology Stack
| Core framework | Python, pip-packaged | Installable, reusable across projects without duplication |
| Methodology | Graduated adversarial pressure | Reports a threshold, not just pass/fail |
Real Design Trade-off, Documented
BCT originally existed as near-duplicate logic embedded in three different projects. Extracting it into a standalone package meant real refactoring work, accepted because three slowly-diverging copies was judged a bigger long-term risk.
Performance, Measured Not Estimated
- Reusable, identical logic now runs across every project needing this testing, via one installable package
See BCT working
A guided walkthrough, one-on-one. We usually reply within one business day.