OWASP top 10 guardrails: what do AI teams still miss?

TL;DR: AI safety policies often fail when organisations rely on static guardrails alone, because prompt injection, excessive agency, sensitive data exposure, and unbounded consumption can still bypass them, according to ActiveFence. The practical issue is governance, not just filtering: teams need continuous red teaming, scoped permissions, and policy controls that match real AI usage patterns.

NHIMG editorial — based on content published by ActiveFence: Aligning AI Safety and Security Policies with the OWASP LLM Top Ten

By the numbers:

• 64% of businesses worry about the integrity of AI systems, and 57% cite trustworthiness as a top challenge.
• 77% of global CEOs are concerned about AI cybersecurity risks.

Questions worth separating out

Q: How should security teams govern AI models that can call tools and access data?

A: Security teams should govern AI models as non-human identities with named owners, limited scope, short-lived credentials, and continuous authorization.

Q: Why do AI guardrails fail when models are connected to business systems?

A: Guardrails fail when they assume the main risk is unsafe language rather than delegated authority.

Q: What do security teams get wrong about governing AI agents?

A: They often treat agents like another automation layer instead of governed non-human actors with their own access paths.

Practitioner guidance

• Separate AI safety policy from identity authority Assign clear owners for prompt safety, tool permissions, secrets handling, and workflow approvals so one team is not pretending to manage the entire risk surface.
• Scope every AI connector and tool call Limit which systems an AI can query, write to, or trigger, and require explicit session boundaries for anything that touches sensitive data or operational actions.
• Test guardrails with adversarial red teaming Include prompt injection, data exfiltration, excessive agency, and output misuse scenarios in recurring testing so policy updates reflect actual abuse paths.

What's in the full article

ActiveFence's full blog covers the operational detail this post intentionally leaves for the source:

• The per-risk policy mappings for each OWASP LLM threat category, including which guardrails ActiveFence associates with prompt injection, data leakage, and excessive agency.
• The concrete examples used to illustrate how each OWASP Top Ten risk can show up in production AI applications and agent workflows.
• The continuous red teaming workflow that ActiveFence describes for finding gaps static policies miss.
• The article's implementation framing for custom policy design across different organisational risk profiles.

👉 Read ActiveFence's analysis of OWASP LLM Top Ten guardrails and AI safety policies →

OWASP top 10 guardrails: what do AI teams still miss?

Explore further

View Full Forum →  |  NHI Foundation Course →