Agentic Skills
Overview
Agentic Skills are callable capabilities exposed by AI agents, each skill represents an action your agents can perform, mapped to underlying APIs or system operations. As skills become a primary attack vector for credential theft, prompt injection, and data exfiltration, understanding and securing your skill footprint is critical.
Akto discovers skills across your entire environment, analyzes each skill file for security risks, and enforces controls to prevent malicious skill execution.
You can view skills by navigating through the Akto Atlas interface:
Akto Atlas → Agentic AI Discovery → Agentic Assets → Skills tab
The Skills table shows endpoint coverage, sensitive data indicators, and last observed activity for each skill. The visibility into usage patterns helps you make informed enforcement decisions.
Skill Discovery Updates
Akto continuously identifies skills configured in your agents.
New skills are discovered every 6 hours. Regular updates keep your inventory aligned with agent configuration changes.
Discovery Sources
Akto discovers skill files across every surface where AI agents are configured and run. Discovery covers:
Local developer workstations — skill files used by IDE-based agents such as Cursor, VS Code, and Claude Code
Source code repositories — skill files committed to GitHub, GitLab, and Bitbucket, including nested directories such as
.claude/skills/CI/CD pipelines — skill files present in pipeline configurations and build environments
MCP servers and AI agent runtimes — skills loaded and exposed by locally running or networked MCP servers
Local developer skill directories — skills loaded from developer-managed local paths outside of version control
Third-party registries and marketplaces — skills pulled in from external sources or community registries
Skill Inventory
Akto generates and maintains a continuous inventory of all skills currently in use across your environment. The inventory captures skill source, hosting location, associated agents, and last observed activity — giving you a complete, up-to-date picture of your skill footprint.
Malicious Skill Detection
When a skill is discovered, Akto analyzes its content for security risks. Each skill is evaluated and assigned a risk score, helping you quickly surface high-risk capabilities and take action before they are exploited. Risk scores are visible across the Skills, Users & Devices, and Agentic Assets views.
Akto detects risks across three risk categories:
Sensitive Data Exposure — hardcoded credentials and sensitive identifiers in skill content
Malicious Instructions & Prompt Injection — injected or adversarial directives designed to override agent behavior
Embedded Code & Script Risks — unsafe execution logic, reverse shells, and dependency vulnerabilities
Analysis runs at discovery time and on every subsequent change detected. For detailed threat context and examples of each risk category, see OWASP Agentic Skills Top 10.
Runtime Enforcement
At runtime, if an injected instruction surfaces during execution, the PromptInjection Guardrail in Agent Guard enforces blocking. See Agent Guard → PromptInjection Guardrail for detection logic and configuration.
Auditing Skills
Once you identify skills that are malicious, over-privileged, or pose a risk to your environment, Akto gives you fine-grained controls to block them at the device or agent level.
Block Skills by Device
You can block a skill directly from the Skills view by selecting the devices where enforcement should apply. Device selection automatically includes all agents running on the selected endpoints.
This approach helps you restrict a capability only where required while allowing normal operation elsewhere.
Steps to Block a Skill for Selected Devices
Navigate the Skills tab under Agentic Assets.
Select the skill that you want to restrict and open its skill details view.
Select the endpoint IDs (devices) where you want to block the skill.
Click Block skill from the action bar at the bottom.
Confirm the blocking action.
Expected Result After Blocking
Akto prevents execution of the selected skill on all chosen devices. All agents running on the selected endpoint IDs lose access to the blocked skill. Agents on other devices continue to use the skill without interruption.
Block Skills by Agent
You can block skills for a specific agent within a specific device using the Users and Devices view. Agent-level enforcement allows you to apply tighter controls without affecting other agents on the same device.
This approach supports environments where different agents require different access levels.
Steps to Block Skills for a Specific Agent
Navigate to Users and Devices.
Select the device where you want to enforce the restriction.
Open the list of agents associated with the selected device and select the agent.
Choose the skills that you want to block.
Click Block Skills from the action bar.
Confirm the blocking action.
Expected Result After Blocking
Akto blocks execution of the selected skills only for the chosen agent on the selected device. Other agents on the same device retain access unless explicitly restricted. Agents on other devices continue to operate without any impact.
Akto Guardrails
Akto monitors agent behavior at runtime to detect dangerous actions triggered by skill execution and enforce policy-based restrictions in real time through agentic guardrails.
What Akto Monitors
Agent tool usage — file access, shell commands, and network calls initiated by skill execution
Skill-triggered system actions — system-level operations spawned directly by a skill
Data movement — files, environment variables, or sensitive content leaving the endpoint
Outbound network activity — destinations contacted during or after skill execution
What Akto Detects
Data exfiltration: environment variable leakage, credential forwarding, or file uploads to external destinations
Unauthorized filesystem activity: access to sensitive paths such as
.envfiles or credential storesAnomalous tool invocation patterns: unusual sequences or frequencies of tool calls that indicate abuse
Pre-authorized dangerous tools configured within skill definitions
Enforcement
Akto enforces controls in real time without requiring manual intervention:
Block dangerous commands immediately upon detection
Restrict filesystem access to sensitive paths based on policy
Limit outbound destinations to approved endpoints only
Integrations
EDR / XDR platforms — Akto correlates skill-triggered events with endpoint telemetry for unified threat visibility
Runtime agent gateways and proxies — Akto integrates with agent middleware to enforce controls at the execution layer
Governance & Compliance
Skill Allowlists & Approval Workflows
Akto gives you controls to govern which skills can be introduced and activated across your environment.
Maintain approved skill registries and allowlists to define which skills are permitted
Require approval workflows before new skills can be introduced to a team or project
Enforce code review gates so skill files go through review before activation
Apply policy gating to prevent skills from being activated without explicit authorization
Usage Tracking & Compliance Reporting
Track skill usage across teams, projects, and devices
Generate compliance reports covering skill inventory and usage history
Align skill governance with security and regulatory requirements
Detection Coverage
Akto measures detection effectiveness across four key risk categories for skills:
Prompt Injection
Detection rate for injected instructions within skill files
Malicious Scripts
Detection rate for dangerous execution logic embedded in or linked from skills
Credential Leakage
Detection rate for hardcoded secrets and sensitive identifiers in skill content
Supply Chain Attacks
Detection rate for compromised or malicious skills introduced via third-party registries or repositories
Detection rates are measured against a benchmark set of known-malicious skill samples covering each risk category. Methodology details are available on request.
Continue with Guardrails
Agentic Skills shows you what your agents can do. Guardrails control what they are allowed to do.
You can:
Map skills to guardrail policies
Enforce restrictions on sensitive actions
Monitor violations
Refer to:
Last updated