Skip to Content
Scan your AI agents for free·npx -y @inkog-io/cli scan .·Get Started →
ComplianceOWASP LLM Top 10

OWASP LLM Top 10 (2026)

Every Inkog finding carries the OWASP GenAI LLM Top 10 2026 category id its rule declares (for example LLM10:2026) in compliance_mapping.owasp_items, and the OWASP Agentic Top 10 2026 id in compliance_mapping.owasp_agentic_threats (CLI JSON, REST API, MCP). The matrix below is generated from that same rule metadata, so it lists exactly what a scan can report. It is the single source for OWASP coverage claims on this site.

Read the Status column: a category counts as detected only when at least one rule stamps its id. Rules that also map to the classic OWASP Web Top 10 keep that id in owasp_web.

Coverage Matrix

OWASP GenAI LLM Top 10 2026. Generated from the rule engine’s metadata; do not edit by hand.

CategoryStatusRules that stamp this idNotes
LLM01: Prompt InjectionActive detectionprompt_injection, prompt_template, skill_tool_poisoningDirect and template-based prompt construction from untrusted input; poisoned skill and tool descriptions.
LLM02: Sensitive Information DisclosureActive detectioncross_tenant, hardcoded_credentials, logging_sensitive_data, pii_filter_wiring, skill_cross_origin, unsafe_env_accessHardcoded credentials, unsafe environment access, sensitive data in logs, missing PII filters, cross-tenant leakage, cross-origin skill data.
LLM03: Excessive AgencyActive detectionexcessive_permissions, goal_conflict_detection, missing_authz, missing_oversight, skill_autonomy_abuse, skill_excessive_permissions, token_validationMissing authorization, missing human oversight, excessive permissions, missing token validation, goal conflicts, autonomous skill abuse.
LLM04: Supply ChainActive detectionskill_supply_chain, supply_chain, unsafe_deserializationUnpinned or untrusted model, tool, package, and skill sources; unsafe deserialization of model or tool artifacts.
LLM05: Data and Model PoisoningActive detectiondata_poisoningUntrusted content written into retrieval stores or memory.
LLM06: Unbounded ConsumptionActive detectioncontext_exhaustion, infinite_loop, missing_rate_limits, recursive_tool_calling, token_bombingInfinite loops, recursive tool calling, token bombing, missing rate limits, context exhaustion.
LLM07: MisinformationActive detectionoverrelianceOverreliance on unverified model output.
LLM08: Hidden Context ExposureActive detectionsystem_prompt_leakSystem prompt and hidden context reachable from logs or user-facing output.
LLM09: Vector and Embedding WeaknessesActive detectionrag_overfetchingRetrieval over-fetching without tenant or scope filters.
LLM10: Improper Output HandlingActive detectioncommand_injection, exec_eval, output_validation, path_traversal, sql_injection, ssrfModel output reaching SQL, shell, eval, file paths, or HTTP targets without validation.

Status definitions

  • Active detection — At least one static-analysis rule stamps this category id on findings. Findings appear under the default balanced policy unless the rule is governance-tier.
  • Governance only — Only governance-layer findings reference this category. They appear under the governance, eu-ai-act, and comprehensive policies, not under balanced or low-noise.
  • Not mapped — No rule currently stamps this category id. Related rules exist but carry a different identifier; a scan will not report this category.
  • Out of scope — Not detectable by static analysis of agent code (for example runtime or infrastructure controls). Inkog does not report this category.

LLM01: Prompt Injection

Status: Active detection. Direct and template-based prompt construction from untrusted input; poisoned skill and tool descriptions.

Rules:

RuleSeverityTierCWE
prompt_injection — Prompt Injection Attack in System PromptHIGHrisk_patternCWE-74, CWE-94, CWE-200
prompt_template — Unvalidated Template Variables in System PromptHIGHrisk_patternCWE-94, CWE-74
skill_tool_poisoning — Tool Description PoisoningCRITICALvulnerabilityCWE-94, CWE-1021

LLM02: Sensitive Information Disclosure

Status: Active detection. Hardcoded credentials, unsafe environment access, sensitive data in logs, missing PII filters, cross-tenant leakage, cross-origin skill data.

Rules:

RuleSeverityTierCWE
cross_tenant — Potential Cross-Tenant Data LeakageCRITICALvulnerabilityCWE-200, CWE-639
hardcoded_credentials — Hardcoded Credentials in System PromptCRITICALvulnerabilityCWE-798, CWE-259
logging_sensitive_data — Sensitive Data in Prompts May Be LoggedMEDIUMrisk_patternCWE-532, CWE-117
pii_filter_wiring — Missing PII Filter Before External CommunicationHIGHvulnerabilityCWE-359, CWE-532
skill_cross_origin — Cross-Origin Tool EscalationHIGHrisk_patternCWE-346, CWE-441
unsafe_env_access — Unsafe Environment Variable AccessHIGHrisk_patternCWE-15, CWE-526

LLM03: Excessive Agency

Status: Active detection. Missing authorization, missing human oversight, excessive permissions, missing token validation, goal conflicts, autonomous skill abuse.

Rules:

RuleSeverityTierCWE
excessive_permissions — Excessive Tool PermissionsMEDIUMrisk_patternCWE-250, CWE-269
goal_conflict_detection — Potential Agent Goal ConflictMEDIUMrisk_patternCWE-362, CWE-667
missing_authz — Missing Authorization Check Before Tool ExecutionHIGHrisk_patternCWE-862
missing_oversight — Agent Executes High-Risk Actions Without Human ApprovalHIGHrisk_patternCWE-862, CWE-1113
skill_autonomy_abuse — Autonomy Abuse in SkillHIGHrisk_patternCWE-862, CWE-770
skill_excessive_permissions — Excessive Permissions in SkillHIGHrisk_patternCWE-250, CWE-269
token_validation — Missing Token Validation Before Sensitive OperationHIGHvulnerabilityCWE-287, CWE-347, CWE-306

LLM04: Supply Chain

Status: Active detection. Unpinned or untrusted model, tool, package, and skill sources; unsafe deserialization of model or tool artifacts.

Rules:

RuleSeverityTierCWE
skill_supply_chain — Skill Package Supply Chain RiskHIGHrisk_patternCWE-829, CWE-494, CWE-1104
supply_chain — Supply Chain Vulnerability in AgentHIGHrisk_patternCWE-829, CWE-494, CWE-1104
unsafe_deserialization — Unsafe Deserialization of LLM/Agent OutputHIGHrisk_patternCWE-502, CWE-915

LLM05: Data and Model Poisoning

Status: Active detection. Untrusted content written into retrieval stores or memory.

Rules:

RuleSeverityTierCWE
data_poisoning — Data/Model Poisoning RiskHIGHrisk_patternCWE-1039, CWE-471, CWE-472

LLM06: Unbounded Consumption

Status: Active detection. Infinite loops, recursive tool calling, token bombing, missing rate limits, context exhaustion.

Rules:

RuleSeverityTierCWE
context_exhaustion — Unbounded Context Window AccumulationMEDIUMrisk_patternCWE-770, CWE-400
infinite_loop — Unbounded Loop in Agentic SystemHIGHrisk_patternCWE-835, CWE-400
missing_rate_limits — Missing Rate Limits on LLM CallsLOWhardeningCWE-770, CWE-799
recursive_tool_calling — Recursive Tool Calling Without Cycle GuardMEDIUMrisk_patternCWE-674, CWE-400
token_bombing — Token Bombing — Unbounded Agent Loop Without Termination GuardHIGHrisk_patternCWE-770, CWE-834

LLM07: Misinformation

Status: Active detection. Overreliance on unverified model output.

Rules:

RuleSeverityTierCWE
overreliance — Overreliance on LLM Output for High-Risk DecisionsMEDIUMrisk_patternCWE-345, CWE-807, CWE-1113

LLM08: Hidden Context Exposure

Status: Active detection. System prompt and hidden context reachable from logs or user-facing output.

Rules:

RuleSeverityTierCWE
system_prompt_leak — System Prompt Leakage (OWASP LLM08:2026)HIGHrisk_patternCWE-209, CWE-532

LLM09: Vector and Embedding Weaknesses

Status: Active detection. Retrieval over-fetching without tenant or scope filters.

Rules:

RuleSeverityTierCWE
rag_overfetching — RAG Over-Fetching - Unbounded Document RetrievalLOWhardeningCWE-400, CWE-770

LLM10: Improper Output Handling

Status: Active detection. Model output reaching SQL, shell, eval, file paths, or HTTP targets without validation.

Rules:

RuleSeverityTierCWE
command_injection — OS Command InjectionCRITICALvulnerabilityCWE-78
exec_eval — Unvalidated Code Execution in Agent ToolHIGHrisk_patternCWE-94, CWE-95, CWE-78
output_validation — Unvalidated LLM Output UsageMEDIUMrisk_patternCWE-20, CWE-116
path_traversal — Path Traversal VulnerabilityHIGHrisk_patternCWE-22, CWE-73, CWE-434
sql_injection — SQL Injection via LLM-Generated QueryHIGHrisk_patternCWE-89, CWE-564
ssrf — SSRF via LLM-Generated URL in Agent ToolHIGHvulnerabilityCWE-918

OWASP Top 10 for Agentic Applications (2026)

OWASP Top 10 for Agentic Applications 2026. Generated from the rule engine’s metadata; do not edit by hand.

CategoryStatusRules that stamp this id
ASI01: Agent Goal HijackActive detectiongoal_conflict_detection, prompt_injection, prompt_template, skill_tool_poisoning
ASI02: Tool Misuse and ExploitationActive detectionoutput_validation, path_traversal, sql_injection, ssrf
ASI03: Identity and Privilege AbuseActive detectioncross_tenant, excessive_permissions, hardcoded_credentials, missing_authz, skill_excessive_permissions, token_validation, unsafe_env_access
ASI04: Agentic Supply Chain VulnerabilitiesActive detectionskill_supply_chain, supply_chain
ASI05: Unexpected Code ExecutionActive detectioncommand_injection, exec_eval, unsafe_deserialization
ASI06: Memory and Context PoisoningActive detectiondata_poisoning, rag_overfetching
ASI07: Insecure Inter-Agent CommunicationActive detectionskill_cross_origin, unsigned_messages
ASI08: Cascading FailuresActive detectioncontext_exhaustion, infinite_loop, missing_rate_limits, recursive_tool_calling, token_bombing
ASI09: Human-Agent Trust ExploitationActive detectionoverreliance, skill_autonomy_abuse
ASI10: Rogue AgentsActive detectionmissing_oversight

Reading OWASP ids in scan output

inkog -path ./agents -output json \ | jq '[.server_findings[] | {id: .pattern_id, owasp: .compliance_mapping.owasp_items}]'

Findings from governance checks (for example missing_human_oversight) only appear under the governance, eu-ai-act, and comprehensive policies. See Security Policies for what each policy filters.

Resources

Last updated on