MediumResource Exhaustion●Realized
Unbounded consumption / denial-of-wallet
Security & RobustnessDescription
Uncontrolled or adversarial usage drives runaway inference cost or service denial.
Example scenario
An attacker scripts expensive long-context queries, spiking the bank's inference bill and degrading service.
Real-world evidence●Realized
The PyTorch torchtriton supply-chain attack (December 2022) introduced malicious code via a PyPI dependency used in AI pipelines, affecting real developer environments. Separately, researchers and security firms documented malicious model files uploaded to Hugging Face Hub in 2023–2024 that executed arbitrary code on load, constituting confirmed production supply-chain compromises in AI infrastructure.
Primary mitigations
- Rate/quota limits
- token budgets
- cost alerts
- abuse detection
- autoscaling guards.
Detection signals
Cost-spike alerts; per-user usage monitoring; abuse-pattern detection.
Mitigating controls
9 Non-agentic controls
ZYC-SEC-001Non-Agentic
Adversarial Testing
Security
ZYC-ROB-001Non-Agentic
Out-of-Distribution Detection
Robustness
ZYC-SCV-001Non-Agentic
Third-Party component risk
Supply Chain Vulnerabilities
ZYC-SYSG-003Non-Agentic
Inference Cost Stability
System Governance
ZYC-SYSG-004Non-Agentic
Latency SLA
System Governance
ZYC-SEC-002Non-Agentic
Fail-Safe Design
Security
ZYC-OREL-003Non-Agentic
Model Denial of Service (DoS)
Over-reliance
ZYC-OREL-004Non-Agentic
Model Denial of Service (DoS)
Over-reliance
ZYC-OREL-005Non-Agentic
Model Denial of Service (DoS)
Over-reliance