OWASP Top 10 for LLM Applications 2025
The OWASP Top 10 for LLM Applications 2025 names ten risks of software that uses a language model. Each entry below lists the Keryn rules that carry it in their metadata (rules 0.8.0) and the edges those rules inspect. The mapping is a reference for people who work in the framework.
This is an edge map, not a score. A count says how many rules look at a class of weakness and how sure they are, not how safe a system is; a class with few rules may simply need few, and no number here means a framework is satisfied. Confidence: P0 exploitable as written, P1 dangerous in context, P2 heuristic or posture. Edges are the relationships the rules inspect (reaches, calls, reads, writes, sends, grants, depends_on, trusts).
Coverage matrix
Section titled “Coverage matrix”| Entry | Rules | P0 | P1 | P2 | Edges |
|---|---|---|---|---|---|
| LLM01 Prompt Injection | 154 | 4 | 88 | 62 | reaches 63 · calls 22 · writes 12 · sends 3 · grants 34 · depends_on 8 · trusts 12 |
| LLM02 Sensitive Information Disclosure | 290 | 17 | 170 | 103 | reaches 39 · calls 13 · reads 31 · writes 17 · sends 74 · grants 106 · depends_on 4 · trusts 5 · none 1 |
| LLM03 Supply Chain | 196 | 20 | 115 | 61 | reaches 14 · calls 25 · reads 8 · sends 19 · grants 26 · depends_on 103 · trusts 1 |
| LLM04 Data and Model Poisoning | 3 | 0 | 1 | 2 | calls 2 · depends_on 1 |
| LLM05 Improper Output Handling | 399 | 30 | 257 | 112 | reaches 32 · calls 268 · reads 15 · writes 13 · sends 6 · grants 22 · depends_on 24 · trusts 19 |
| LLM06 Excessive Agency | 588 | 61 | 314 | 213 | reaches 61 · calls 187 · reads 32 · writes 26 · sends 31 · grants 205 · depends_on 19 · trusts 27 |
| LLM07 System Prompt Leakage | 78 | 3 | 42 | 33 | reaches 14 · calls 6 · reads 7 · writes 11 · sends 21 · grants 9 · depends_on 3 · trusts 4 · none 3 |
| LLM08 Vector and Embedding Weaknesses | 109 | 0 | 48 | 61 | reaches 6 · calls 20 · reads 1 · writes 23 · grants 58 · depends_on 1 |
| LLM09 Misinformation | 30 | 1 | 23 | 6 | calls 8 · reads 6 · grants 3 · depends_on 3 · trusts 7 · none 3 |
| LLM10 Unbounded Consumption | 99 | 2 | 41 | 56 | reaches 13 · calls 10 · writes 4 · sends 2 · grants 67 · depends_on 1 · none 2 |
LLM01 Prompt Injection
Section titled “LLM01 Prompt Injection”154 rules · confidence P0 4, P1 88, P2 62 · severity error 41, warning 103, info 10
Edges inspected: reaches 63 · calls 22 · writes 12 · sends 3 · grants 34 · depends_on 8 · trusts 12
30 of 154 rules, surest first
- seceng.agentrules.secrets-to-url-command
- seceng.ciagent.agent-pr-head-checkout-on-pull-request-target
- seceng.mcp.py.tool-description-hidden-unicode
- seceng.mcp.ts.tool-description-hidden-unicode
- keryn.canon.go.validate-before-decode
- keryn.canon.php.validate-before-decode
- keryn.canon.py.validate-before-decode
- keryn.canon.ts.validate-before-decode
- seceng.agentrules.base64-decode-exec
- seceng.agentrules.bidi-override-character
- seceng.agentrules.invisible-unicode-character
- seceng.aiconfig.bidi-override-character
- seceng.aiconfig.escaped-invisible-unicode
- seceng.aiconfig.invisible-unicode-character
- seceng.asi01.agent-instructions-loaded-at-runtime-python
- seceng.asi01.agent-instructions-loaded-at-runtime-ts
- seceng.asi01.system-prompt-from-retrieved-or-tool-content-python
- seceng.asi01.system-prompt-from-retrieved-or-tool-content-ts
- seceng.asi06.langgraph-langchain-memory-write-from-untrusted-python
- seceng.asi06.memory-read-into-system-prompt-python
- seceng.asi06.memory-read-into-system-prompt-ts
- seceng.asi07.peer-agent-message-as-instructions-python
- seceng.asi07.peer-agent-message-as-instructions-ts
- seceng.ciagent.agent-prompt-interpolates-event-text
- seceng.ciagent.agent-step-on-pull-request-target
- seceng.llm01.prompt-injection-taint
- seceng.llm01.taint.tool-result-to-tool-args-python
- seceng.mcp.py.tool-description-instructions
- seceng.mcp.py.tool-docstring-instructions
- seceng.mcp.ts.tool-description-instructions
LLM02 Sensitive Information Disclosure
Section titled “LLM02 Sensitive Information Disclosure”290 rules · confidence P0 17, P1 170, P2 103 · severity error 82, warning 172, info 36
Edges inspected: reaches 39 · calls 13 · reads 31 · writes 17 · sends 74 · grants 106 · depends_on 4 · trusts 5 · none 1
30 of 290 rules, surest first
- keryn.skillcred.browser-credential-store-read
- keryn.skillcred.cloud-metadata-endpoint-access
- keryn.skillcred.macos-keychain-secret-read
- keryn.skillcred.os-credential-database-access
- keryn.skillcred.secret-manager-or-keyring-export
- keryn.skillcred.ssh-key-material-read
- seceng.agentcfg.mcp-docker-sensitive-mount
- seceng.agentcfg.mcp-env-inline-secret-known-prefix
- seceng.agentrules.secrets-to-url-command
- seceng.iacai.gcp-ai-role-granted-to-public
- seceng.mcp.py.oauth-client-secret-hardcoded
- seceng.mcp.ts.oauth-client-secret-hardcoded
- seceng.eval.mlflow-log-raw-prompt
- seceng.eval.public-dashboard-no-auth
- seceng.eval.weave-trace-without-redaction
- seceng.llm-frameworks.llamaindex-insecure-storage
- seceng.llm02-system-prompt-in-client
- keryn.browseragent.browser-launch-disable-web-security-js
- keryn.browseragent.browser-launch-disable-web-security-py
- keryn.egress.cs.secret-to-side-channel
- keryn.egress.go.secret-to-side-channel
- keryn.egress.java.secret-to-side-channel
- keryn.egress.php.secret-to-side-channel
- keryn.egress.py.secret-to-side-channel
- keryn.egress.ts.secret-to-side-channel
- keryn.fastapi.cors-wildcard-with-credentials
- keryn.general.python.paramiko-auto-add-host-key
- keryn.general.python.ssl-insecure-protocol
- keryn.keyliteral.anthropic
- keryn.keyliteral.elevenlabs
LLM03 Supply Chain
Section titled “LLM03 Supply Chain”196 rules · confidence P0 20, P1 115, P2 61 · severity error 71, warning 112, info 13
Edges inspected: reaches 14 · calls 25 · reads 8 · sends 19 · grants 26 · depends_on 103 · trusts 1
30 of 196 rules, surest first
- keryn.agenthooks.copilot-hook-pipes-download-to-shell
- keryn.agenthooks.cursor-hook-pipes-download-to-shell
- keryn.agenthooks.gemini-hook-pipes-download-to-shell
- keryn.agenthooks.windsurf-hook-pipes-download-to-shell
- keryn.modelload.cloudpickle-or-dill-loads-py
- keryn.modelload.hf-hub-download-result-unpickled-py
- keryn.modelload.keras-load-model-safe-mode-disabled-py
- keryn.modelload.native-extension-library-load-py
- keryn.modelload.numpy-load-allow-pickle-py
- keryn.modelload.torch-hub-load-state-dict-nonliteral-url-py
- keryn.modelload.torch-hub-load-trust-repo-nonliteral-py
- keryn.modelload.yaml-load-unsafe-loader-py
- keryn.skillexec.reverse-shell-bash-devtcp
- keryn.skillexec.reverse-shell-netcat
- keryn.skillexec.reverse-shell-python
- keryn.skillexec.reverse-shell-script-langs
- keryn.skillexec.reverse-shell-socat
- seceng.agentcfg.hook-pipes-to-shell
- seceng.mcp.py.stdio-server-command-from-request
- seceng.mcp.ts.stdio-client-transport-command-from-request
- keryn.general.csharp.insecure-deserialization-javascriptserializer-typeresolver
- keryn.general.js.invisible-hangul-filler
- keryn.general.js.trojan-source-bidi-control
- keryn.keyliteral.huggingface
- keryn.modelload.pandas-read-pickle-py
- keryn.modelload.request-path-to-model-loader-python
- keryn.modelload.torch-package-importer-py
- keryn.skillexec.base64-exec-script-langs
- keryn.skillexec.decrypt-pipe-to-interpreter
- keryn.skillexec.download-chmod-run
LLM04 Data and Model Poisoning
Section titled “LLM04 Data and Model Poisoning”3 rules · confidence P0 0, P1 1, P2 2 · severity error 0, warning 2, info 1
Edges inspected: calls 2 · depends_on 1
3 rules
LLM05 Improper Output Handling
Section titled “LLM05 Improper Output Handling”399 rules · confidence P0 30, P1 257, P2 112 · severity error 116, warning 248, info 35
Edges inspected: reaches 32 · calls 268 · reads 15 · writes 13 · sends 6 · grants 22 · depends_on 24 · trusts 19
30 of 399 rules, surest first
- keryn.browseragent.model-output-to-page-script-js
- keryn.browseragent.model-output-to-page-script-py
- keryn.general.ruby.command-injection-request
- keryn.general.ruby.erb-template-from-request
- keryn.general.ruby.eval-request
- keryn.general.ruby.unsafe-deserialization-request
- keryn.rustagent.model-output-to-process-command-rs
- keryn.rustagent.model-output-to-sql-rs
- keryn.rustagent.model-output-to-webview-eval-rs
- seceng.llm05.taint.model-output-to-async-subprocess-python
- seceng.llm05.taint.model-output-to-browser-eval-js
- seceng.llm05.taint.model-output-to-child-process-js
- seceng.llm05.taint.model-output-to-code-interpreter-tool-python
- seceng.llm05.taint.model-output-to-deserialization-python
- seceng.llm05.taint.model-output-to-dynamic-import-python
- seceng.llm05.taint.model-output-to-eval-js
- seceng.llm05.taint.model-output-to-eval-python
- seceng.llm05.taint.model-output-to-mcp-stdio-command-js
- seceng.llm05.taint.model-output-to-os-command-python
- seceng.llm05.taint.model-output-to-remote-command-python
- seceng.llm05.taint.model-output-to-require-import-js
- seceng.llm05.taint.model-output-to-shell-quote-missing-js
- seceng.llm05.taint.model-output-to-sql-js
- seceng.llm05.taint.model-output-to-sql-python
- seceng.llm05.taint.model-output-to-template-engine-js
- seceng.llm05.taint.model-output-to-template-python
- seceng.mcp.csharp.tool-sql-command-interpolated
- seceng.mcp.go.tool-sql-string-built
- seceng.mcp.java.tool-jdbc-statement-concat
- seceng.mcp.rust.tool-sqlx-query-format
LLM06 Excessive Agency
Section titled “LLM06 Excessive Agency”588 rules · confidence P0 61, P1 314, P2 213 · severity error 234, warning 315, info 39
Edges inspected: reaches 61 · calls 187 · reads 32 · writes 26 · sends 31 · grants 205 · depends_on 19 · trusts 27
30 of 588 rules, surest first
- keryn.agenthooks.copilot-hook-pipes-download-to-shell
- keryn.agenthooks.copilot-hook-reverse-shell
- keryn.agenthooks.cursor-hook-pipes-download-to-shell
- keryn.agenthooks.cursor-hook-reverse-shell
- keryn.agenthooks.gemini-hook-pipes-download-to-shell
- keryn.agenthooks.gemini-hook-reverse-shell
- keryn.agenthooks.windsurf-hook-pipes-download-to-shell
- keryn.agenthooks.windsurf-hook-reverse-shell
- keryn.cryptoclaim.cs.jwt-verification-disabled
- keryn.cryptoclaim.go.jwt-verification-disabled
- keryn.cryptoclaim.java.jwt-verification-disabled
- keryn.cryptoclaim.kotlin.jwt-verification-disabled
- keryn.cryptoclaim.php.jwt-verification-disabled
- keryn.cryptoclaim.py.jwt-verification-disabled
- keryn.cryptoclaim.ts.jwt-verification-disabled
- keryn.outbound.model-output-as-payment-amount-js
- keryn.outbound.model-output-as-payment-amount-py
- keryn.rustagent.model-output-to-process-command-rs
- keryn.skillcred.browser-credential-store-read
- keryn.skillcred.cloud-metadata-endpoint-access
- keryn.skillcred.container-escape-from-skill
- keryn.skillcred.macos-keychain-secret-read
- keryn.skillcred.os-credential-database-access
- keryn.skillcred.secret-manager-or-keyring-export
- keryn.skillcred.ssh-key-material-read
- keryn.skillexec.reverse-shell-bash-devtcp
- keryn.skillexec.reverse-shell-netcat
- keryn.skillexec.reverse-shell-python
- keryn.skillexec.reverse-shell-script-langs
- keryn.skillexec.reverse-shell-socat
LLM07 System Prompt Leakage
Section titled “LLM07 System Prompt Leakage”78 rules · confidence P0 3, P1 42, P2 33 · severity error 1, warning 43, info 34
Edges inspected: reaches 14 · calls 6 · reads 7 · writes 11 · sends 21 · grants 9 · depends_on 3 · trusts 4 · none 3
30 of 78 rules, surest first
- seceng.eval.public-dashboard-no-auth
- seceng.llm-frameworks.dspy-signature-leakage
- seceng.llm02-system-prompt-in-client
- seceng.llm07.persona-tampering
- keryn.agenthooks.copilot-prompt-hook-posts-to-network
- keryn.agenthooks.cursor-prompt-hook-posts-to-network
- keryn.agenthooks.gemini-prompt-hook-posts-to-network
- keryn.agenthooks.windsurf-prompt-hook-posts-to-network
- seceng.bundle-embedded-system-prompt
- seceng.bundle-tool-schema-exposed
- seceng.edge.vercel-edge-config-secret-or-prompt
- seceng.electron-preload-exposes-ai-tool-api
- seceng.graphql-schema-exposes-prompt-or-tool-types
- seceng.kotlin.system-prompt-in-mobile-client
- seceng.llm02.prompt-message-logging
- seceng.llm02.prompt-message-logging-rust
- seceng.llm07.env-var-prompt-in-response
- seceng.llm07.env-var-prompt-in-response-js
- seceng.llm07.insecure-prompt-storage
- seceng.postgres-prompt-or-secret-column-in-public-view
- seceng.swift.system-prompt-in-mobile-client
- seceng.tauri-ai-prompt-or-key-in-frontend
- seceng.wordpress-transcript-or-prompt-option-storage
- seceng.bundle-source-map-with-ai-secrets-risk
- seceng.cloudai.bedrock-prompt-logging-or-tracing
- seceng.eval.langfuse-raw-input-output-logging
- seceng.forensics-ai-error-swallowed
- seceng.forensics-no-user-tenant-on-ai-event
- seceng.forensics-prompt-hash-missing
- seceng.forensics.ai-action-missing-correlation-id
LLM08 Vector and Embedding Weaknesses
Section titled “LLM08 Vector and Embedding Weaknesses”109 rules · confidence P0 0, P1 48, P2 61 · severity error 6, warning 83, info 20
Edges inspected: reaches 6 · calls 20 · reads 1 · writes 23 · grants 58 · depends_on 1
30 of 109 rules, surest first
- seceng.asi06.langgraph-langchain-memory-write-from-untrusted-python
- seceng.asi06.memory-read-into-system-prompt-python
- seceng.asi06.memory-read-into-system-prompt-ts
- seceng.asi06.memory-shared-user-scope-python
- seceng.asi06.memory-shared-user-scope-ts
- seceng.ragloader.ingest-pipeline-fetch-user-url
- seceng.cloudflare-d1-ai-query-without-tenant-filter
- seceng.cloudflare-kv-ai-memory-without-tenant-prefix
- seceng.cloudflare-r2-public-ai-docs
- seceng.deno.edge-kv-ai-memory-without-tenant-prefix
- seceng.firebase-storage-public-ai-artifact
- seceng.firebase.firestore-ai-collection-read-without-owner-filter
- seceng.llm05.taint.model-output-to-memory-write-js
- seceng.llm08.supabase-match-documents-rpc
- seceng.mongoose.ai-query-without-tenant-filter
- seceng.orm.drizzle-ai-query-without-tenant-filter
- seceng.orm.prisma-ai-findmany-without-tenant-filter
- seceng.orm.sequelize-ai-resource-without-owner
- seceng.orm.sqlalchemy-ai-query-without-tenant-filter
- seceng.orm.typeorm-findone-ai-resource-without-owner
- seceng.postgres-ai-function-dynamic-sql
- seceng.postgres-ai-table-without-rls
- seceng.postgres-security-definer-rag-function
- seceng.redis-url-without-tls
- seceng.supabase-auth-user-not-bound-to-query
- seceng.supabase-storage-public-ai-documents
- seceng.supabase.postgrest-public-table-ai-data
- seceng.supabase.rls-not-enabled-for-ai-table
- seceng.supabase.rls-policy-without-auth-uid
- seceng.supabase.rpc-query-text-directly-from-request
LLM09 Misinformation
Section titled “LLM09 Misinformation”30 rules · confidence P0 1, P1 23, P2 6 · severity error 2, warning 9, info 19
Edges inspected: calls 8 · reads 6 · grants 3 · depends_on 3 · trusts 7 · none 3
30 rules
- seceng.llm09.missing-hallucination-disclaimer
- seceng.asi08.tool-error-swallowed-as-success-python
- seceng.asi08.tool-error-swallowed-as-success-ts
- keryn.vendor.anthropic-content-read-without-stop-reason-check-js
- keryn.vendor.anthropic-content-read-without-stop-reason-check-py
- keryn.vendor.openai-content-read-without-refusal-check-js
- keryn.vendor.openai-content-read-without-refusal-check-py
- seceng.schema-model-score-controls-security-decision
- seceng.dataset.eval-set-from-production-no-redaction
- seceng.dataset.hf-unknown-user-repo-for-finetune
- seceng.dataset.remote-training-data-no-integrity
- seceng.dataset.synthetic-data-leaks-source-records
- seceng.forensics-model-version-missing
- seceng.forensics-no-user-tenant-on-ai-event
- seceng.forensics-prompt-hash-missing
- seceng.forensics.ai-action-missing-correlation-id
- seceng.forensics.tool-call-missing-audit-event
- seceng.github-ai-pr-bot-commits-without-approval
- seceng.gitlab-ai-generated-commit-no-approval
- seceng.llm09.js-ts-rag-response-missing-citation
- seceng.llm09.js-ts-rag-response-missing-confidence
- seceng.llm09.python-rag-response-missing-citation
- seceng.llm09.python-rag-response-missing-confidence
- seceng.react.ai-output-to-codeblock-exec-affordance
- seceng.dataset.hf-dataset-unpinned
- seceng.repoagent.model-output-to-git-command
- seceng.schema.model-json-parse-without-validation-python
- seceng.schema.model-json-parse-without-validation-ts
- seceng.schema-openai-json-mode-without-schema
- seceng.vercelai-generateobject-without-schema
LLM10 Unbounded Consumption
Section titled “LLM10 Unbounded Consumption”99 rules · confidence P0 2, P1 41, P2 56 · severity error 11, warning 60, info 28
Edges inspected: reaches 13 · calls 10 · writes 4 · sends 2 · grants 67 · depends_on 1 · none 2
30 of 99 rules, surest first
- keryn.skillpersist.fork-bomb
- seceng.iacai.gcp-ai-role-granted-to-public
- seceng.asi08.agent-retry-without-cap-python
- seceng.asi08.agent-retry-without-cap-ts
- seceng.asi08.unbounded-agent-fanout-from-model-output-python
- seceng.asi08.unbounded-agent-fanout-from-model-output-ts
- seceng.iacai.azure-openai-network-acls-default-allow
- seceng.iacai.azure-openai-public-network-access-enabled
- seceng.iacai.k8s-ai-service-exposed-loadbalancer-nodeport
- seceng.modelsc.ollama-pull-user-model
- seceng.modelsc.ollama-pull-user-model-ts
- keryn.gaps.unbounded-model-loop
- keryn.general.python.http-request-without-timeout
- keryn.skillpersist.disk-or-memory-fill
- seceng.cloudai.vertex-broad-aiplatform-permissions
- seceng.cloudai.vertex-public-endpoint-or-logging
- seceng.iacai.azure-openai-no-network-acls
- seceng.iacai.vllm-container-no-api-key
- seceng.llm05.taint.model-output-to-regex-python
- seceng.llm10.js-ts-agentic-loop-dos-logic
- seceng.llm10.openai-call-without-max-tokens
- seceng.llm10.python-agentic-loop-dos-logic
- seceng.llm10.unbounded-retry-loop
- seceng.llm10.unbounded-retry-loop-js
- seceng.mongoose-ai-regex-query-from-model-output
- seceng.tsorm-ai-generated-orderby-or-select
- seceng.agentframework.crewai-process-without-rate-or-step-limit
- seceng.cloudai.azure-openai-public-network-access
- seceng.container-compose-ai-port-public-no-auth
- seceng.container.ai-service-loadbalancer-no-auth-hint