AI Security · Runtime · Supply Chain

Postura de segurança das tecnologias de IA

Avaliação dos quinze vetores de ataque de IA, dos nove controles de runtime governance, da descoberta contínua de AI-SPM e da exposição da cadeia de suprimentos de modelos, MCPs e agentes.

Cobertura AI Security

21%

15 vetores avaliados

Cobertura Runtime

32%

9 controles de runtime

Exposição cibernética máx.

10/10

Tecnologias com privilégio excessivo

controle ausente

Lacunas — AI Security (OWASP Top 10 LLM / SANS)

Cobertura do controle no portfólio, do menor para o maior.

Indirect Prompt Injection (0/16)0%

Agent Takeover (0/16)0%

Lateral Movement (0/16)0%

RAG Poisoning (1/16)6%

Hallucination (2/16)13%

Data Exfiltration (2/16)13%

Adversarial Attacks (2/16)13%

Privilege Escalation (2/16)13%

Model Poisoning (3/16)19%

Supply Chain Attack (3/16)19%

Prompt Injection (4/16)25%

Excessive Permissions (5/16)31%

Jailbreak (7/16)44%

Secrets Exposure (8/16)50%

Data Leakage (11/16)69%

Lacunas — AI Runtime Governance

Cobertura do controle no portfólio, do menor para o maior.

Runtime Platform Risk (0/16)0%

Runtime Resilience (1/16)6%

Runtime Enforcement (1/16)6%

Runtime Segregation (3/16)19%

Runtime Policy Controls (3/16)19%

Runtime Observability (4/16)25%

Runtime Security (7/16)44%

Runtime Monitoring (13/16)81%

Runtime Logging (15/16)94%

Exposição de segurança por tecnologia

Tecnologia	Exp. segurança	Security ctrl.	Runtime ctrl.	Supply chain	Prioridade
Agentes autônomos de back-office	10.0	7%	11%	7.0	Auditoria imediata
MCP Gateway corporativo	9.5	13%	22%	9.0	Próximo ciclo
Microsoft 365 Copilot	9.0	27%	33%	4.0	Monitoramento contínuo
Azure OpenAI Service	8.0	33%	44%	7.0	Próximo ciclo
GitHub Copilot / Cursor	8.0	20%	22%	8.0	Monitoramento contínuo
Glean (AI Search corporativa)	8.0	13%	22%	6.0	Monitoramento contínuo
ServiceNow AI Agents	8.0	13%	22%	7.0	Monitoramento contínuo
ChatGPT Enterprise	8.0	20%	11%	6.0	Monitoramento contínuo
UiPath / Automation Anywhere Agentic	8.0	13%	33%	7.0	Próximo ciclo
Amazon Bedrock	7.0	27%	56%	8.0	Monitoramento contínuo
Salesforce Agentforce	7.0	20%	33%	6.0	Monitoramento contínuo
Google Gemini / Vertex AI	7.0	20%	33%	7.0	Monitoramento contínuo
AI Security Posture (Palo Alto / CrowdStrike)	7.0	40%	44%	6.0	Observação
Copiloto de crédito (motor de decisão)	6.5	27%	44%	6.0	Próximo ciclo
Databricks Mosaic AI	6.0	27%	56%	6.0	Monitoramento contínuo
AI Observability (Arize / LangSmith)	6.0	13%	33%	5.0	Observação

Context Engineering

Risco médio do domínio6.7/10

Context Management
Context Lineage
Context Security
Context Classification
Context Leakage
Context Poisoning
Retrieval Context Quality

AI-SPM

Risco médio do domínio6.6/10

Model Discovery
Agent Discovery
MCP Discovery
Attack Surface Mapping
Vulnerability Identification
Security Drift
Exposure Mapping

AI Supply Chain

Risco médio do domínio6.6/10

Open Source Models
Marketplace Models
Third Party Providers
External Dependencies
External APIs
MCP Marketplaces
Agent Marketplaces

Agentes autônomos de back-office

Fornecedores e dependências: LangGraph · MCP marketplace público · APIs transacionais internas

Vetores de ataque cobertos

1/15 · 7%

Prompt Injection
Indirect Prompt Injection
Jailbreak
Hallucination
Data Leakage
Data Exfiltration
RAG Poisoning
Model Poisoning
Adversarial Attacks
Supply Chain Attack
Secrets Exposure
Excessive Permissions
Agent Takeover
Privilege Escalation
Lateral Movement

Runtime governance

1/9 · 11%

Runtime Security
Runtime Observability
Runtime Resilience
Runtime Monitoring
Runtime Logging
Runtime Segregation
Runtime Enforcement
Runtime Policy Controls
Runtime Platform Risk

MCP Gateway corporativo

Fornecedores e dependências: MCP registries públicos · Servidores MCP de terceiros · Conectores comunitários

Vetores de ataque cobertos

2/15 · 13%

Prompt Injection
Indirect Prompt Injection
Jailbreak
Hallucination
Data Leakage
Data Exfiltration
RAG Poisoning
Model Poisoning
Adversarial Attacks
Supply Chain Attack
Secrets Exposure
Excessive Permissions
Agent Takeover
Privilege Escalation
Lateral Movement

Runtime governance

2/9 · 22%

Runtime Security
Runtime Observability
Runtime Resilience
Runtime Monitoring
Runtime Logging
Runtime Segregation
Runtime Enforcement
Runtime Policy Controls
Runtime Platform Risk

Microsoft 365 Copilot

Fornecedores e dependências: OpenAI (via Azure) · Microsoft Graph · Marketplace de agentes

Vetores de ataque cobertos

4/15 · 27%

Prompt Injection
Indirect Prompt Injection
Jailbreak
Hallucination
Data Leakage
Data Exfiltration
RAG Poisoning
Model Poisoning
Adversarial Attacks
Supply Chain Attack
Secrets Exposure
Excessive Permissions
Agent Takeover
Privilege Escalation
Lateral Movement

Runtime governance

3/9 · 33%

Runtime Security
Runtime Observability
Runtime Resilience
Runtime Monitoring
Runtime Logging
Runtime Segregation
Runtime Enforcement
Runtime Policy Controls
Runtime Platform Risk