CI/CD gates
Decide which security findings should block a release.
METHODOLOGY & SCOPE
What an AxiomGate audit inspects.
Scoping areas, the risk taxonomy they map to, and the architecture patterns behind the findings. Your confirmed audit defines which checks and artifacts are included.
TEN REVIEW AREAS / ONE GOVERNANCE STORY
From the first tool call
to the evidence you keep.
Choose the controls and evidence your system needs. These are scoping areas; your confirmed audit defines the checks and artifacts included.
Agent SBOM
Map agents, dependencies, and who delegates to whom.
Runtime Guard
Review policy mediation and the limits of each capability.
MCP review
Inspect tool schemas, exposed authority, and trust boundaries.
STRIDE threats
Turn architecture boundaries into concrete threat scenarios.
Delegation stress tests
Scope adversarial paths and investigate expanding authority.
Trace analysis
Follow delegation hops and policy decisions in supplied traces.
Compliance mapping
Organize technical evidence for your compliance review.
Formal boundaries
Separate model-checked properties from implementation assumptions.
Static security audit
Prioritize actionable findings without executing the submitted code.
Integration, runtime tests, and compliance work need explicit scope. No framework selection implies complete automatic coverage or certification.
RISK TAXONOMY
The risks these checks map to.
Prompt Injection & Goal Hijacking
Untrusted web content, documents, or API responses can redirect an agent’s goals. We inspect how tool authority is constrained independently of model instructions.
Excessive Agency & Delegation Loops
Over-privileged tools and unbounded inter-agent delegation amplify errors. We audit capability attenuation, recursion budgets, and commit barriers.
System Prompt & Credential Leakage
Prompts are not secrets. We inspect DLP boundaries and credential storage so private keys and API tokens are not copied into report artifacts or hosted-model contexts by the intake path.
Risk taxonomy: OWASP Top 10 for LLM Applications (2025 Edition) ↗. Structured static analysis with explicit coverage limits.
Penalty ceilings vary by violation: Article 99 allows up to €35M or 7% of worldwide annual turnover for prohibited practices. This is technical evidence, not legal advice. Read the Act ↗
03 / BOUNDARIES IN PRACTICE
See the decision
before the damage.
Three real-world architecture patterns.
Verifiable code with before and after mediation.
Unconstrained shell execution in CrewAI
ILLUSTRATIVE PATTERN VULNERABLE / AUTHORITY IMPLICIT
HARDENED / EXPLICIT CONTROL