AI Model Risk Depends on What the Model Can Reach

Snyk’s 2026 State of Agentic AI Adoption, Volume III, analyzes anonymized scans of 3,023 enterprise environments and model-specific adversarial testing across 29 attacker goals to argue that model risk scores should guide investigation rather than serve as standalone evidence of exploitation. Among 1,693 environments with detected model identifiers, 74.5% referenced a model with at least one high-risk score and 49.2% referenced a critical-risk model, particularly for personal-data extraction and command execution, although only 37.5% of detected model occurrences could be matched to a score. The report emphasizes that an AI model’s practical risk depends on its deployment context, as agents and MCP servers can connect models to sensitive data, tools, and business actions; organizations with both agent indicators and MCP server-building code showed substantially more high-scoring model references. It recommends verifying which models, agents, servers, tools, and permissions are active, selecting alternative models when scores differ meaningfully by attack type, and relying on application-level safeguards such as least privilege, output validation, approvals, monitoring, and code scanning when weaknesses are common across models. The research also finds that lower-capability models tended to perform worse on several security measures, premium pricing did not consistently correlate with stronger security, open-weight models had higher average personal-data leakage scores in a small sample, and missing training-data lineage remains common, especially for proprietary models.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.