>
    Strata Copilot
    AI Skill Security Rules
    Focus
    Download PDF
    Focus
    1. Home
    2. Prisma AIRS
    3. AI Supply Chain Security
    4. AI Supply Chain Security
    5. AI Skill Security (Preview)
    6. AI Skill Security Rules
    Download PDF
    Prisma AIRS

    AI Skill Security Rules

    Table of Contents

    AI Skill Security Rules

    Learn about the built-in security rules that evaluate uploaded skills for security and governance risks, and how rule settings control scan verdicts.
    Where Can I Use This?What Do I Need?
    • Prisma AIRS (AI Skill Security)
    • No License required for preview
    Each uploaded skill is evaluated against a set of built-in security rules designed to identify behaviors that may introduce security or operational risk. A single Default Security Group contains all built-in security and governance rules evaluated during every scan.
    RuleDescriptionRule TypeDefault State
    Prompt Integrity CheckDetects embedded instructions intended to manipulate or override expected agent behavior.StaticEnabled + Blocking
    Arbitrary Code Execution CheckDetects executable code or patterns capable of performing actions beyond the skill's documented purpose.StaticEnabled + Blocking
    Secrets Disclosure CheckDetects unauthorized access to credentials, API keys, tokens, and other sensitive information.StaticEnabled + Blocking
    Data Exfiltration CheckDetects outbound communication or data transfers that may expose sensitive information.StaticEnabled + Blocking
    Obfuscated Behavior CheckDetects actions performed without appropriate visibility, logging, or user awareness.StaticEnabled + Blocking
    Behavior Integrity CheckDetects modifications to agent configuration, instructions, or persistent state.StaticEnabled + Blocking
    Excessive Permissions CheckDetects autonomous behaviors or capabilities that exceed expected operational boundaries.StaticEnabled + Allowing
    Outbound Network Activity CheckDetects whether the skill contacts the network at runtime.DynamicEnabled + Allowing
    Credential Access CheckDetects whether the skill reads secrets, SSH keys, tokens, or environment material at runtime.DynamicEnabled + Allowing
    Data Exfiltration Staging CheckDetects whether the skill stages or prepares for the transmission of data outside the expected boundary at runtime.DynamicEnabled + Allowing
    Process and Shell Execution CheckDetects whether the skill attempts to spawn shells, reverse shells, or perform privilege escalation at runtime.DynamicEnabled + Allowing
    Filesystem Tampering CheckDetects whether the skill makes sensitive filesystem changes.DynamicEnabled + Allowing
    Defense Evasion CheckDetects whether the skill actively tries to bypass controls on it at runtime.DynamicEnabled + Allowing
    Persistence CheckDetects whether the skill attempts to persist execution or data for future use beyond the scope of its invocation.DynamicEnabled + Allowing
    Each rule is controlled by two toggles:
    • Enabled—When this toggle is off, the rule is not evaluated.
    • Blocking—When this toggle is on, a finding from the rule's detectors produces a Blocked verdict. When it is off, the finding is still recorded but does not affect the verdict (alert only). The Blocking toggle is unavailable while Enabled is off.
    EnabledBlockingBehavior
    OnOnBlock — a finding fails the scan verdict.
    OnOffAlert only — finding recorded, not counted towards scan outcome.
    OffNADisabled — not evaluated.
    Configuration changes apply only to subsequent scans. In-flight and completed scans are not affected. To modify rule settings, see Configure AI Skill Security Rules.

    Static and Dynamic Rules

    AI Skill Security evaluates each skill with two types of rules:
    • Static—An analysis of the static contents of the skill, including parsing any markdown and code files for vulnerabilities and unsafe or deceptive behaviors.
    • Dynamic—An analysis of a skill performed by loading the skill into an agent in a sandboxed environment and observing the actual behavior that the various functions of the skill induce.

    Rule Evaluation States

    Each rule that applies to a scan reports one of the following evaluation states:
    • Failed—The skill violated the rule at least once.
    • Passed—The skill did not violate the rule.
    • Skipped—The rule was not evaluated on the skill. This can happen for skills that have already been classified as safe or unsafe with very high confidence, and therefore the additional dynamic analysis isn't necessary.
    • Error—An error occurred while the rule was being evaluated. Try the analysis again at a later time.

    © 2026 Palo Alto Networks, Inc. All rights reserved.