How InnerWarden covers the ten OWASP agentic risks.
InnerWarden screens every command and tool call an AI agent tries to run, and on Linux the kernel Execution Gate refuses the dangerous ones before they execute. This page maps that protection to the OWASP Top 10 for Agentic Applications 2026 and labels each risk honestly: direct, partial, or supporting.
Direct on the runtime risks, honest about the rest. All ten of ASI01-ASI10 carry a label below.
Coverage matrix
Ten agentic risks, mapped to tested runtime controls.
Strong on the runtime-observable risks, honest about the ones a host guardrail does not fully cover. The label on each card says which.
ASI01
Agent Goal Hijack
Detect
Prompt-injection detection (27 patterns) on the commands, arguments and tool responses routed through the guard; obfuscation flagged before it runs. Injected user input is deny; an instruction injected into a tool result is a review alert, not blocked.
ASI02
Tool Misuse & Exploitation
Direct · kernel (Linux)
check-command denies the dangerous tool calls marked blocking and holds the rest for review (not blocked for an agent); a poisoned tool description is denied; the MCP proxy's loop breaker refuses an identical tool call repeated more than 3 times within 60 s; the armed kernel Execution Gate (Linux) can enforce the exec.
ASI03
Identity & Privilege Abuse
Partial
Reading or overwriting a credential store is denied; searching the filesystem for credentials is review; loosened permissions on a system path are flagged; Pro and Enterprise add privilege-provenance signals. It does not manage the agent's own identity, tokens or delegation.
ASI04
Agentic Supply Chain Vulnerabilities
Name screening · runtime
No SBOM, provenance, signature, version-pinning or registry validation. Community flags an install (pip, pipx, uv, poetry, pipenv, npm, yarn, pnpm, bun), a fetch-and-run (npx, npm exec, pnpm dlx, yarn dlx, bunx, uvx, uv tool run, pipx run) or a download (pip download, pip wheel) whose package name is one edit from a popular PyPI or npm package and is not itself on the shipped list: review, blocked for an agent. Community can also detect a payload attempt; an armed Execution Gate (Pro and Enterprise) can block its execution on Linux.
ASI05
Unexpected Code Execution
Direct · kernel (Linux)
Measured on the shipped benchmark: a download piped to a shell, or saved to a file (or unpacked from a fetched archive, or installed as a fetched package) and then run (curl, wget, BSD fetch, aria2c, axel, lwp-download, a Python, Node, Perl, Ruby or PHP one-liner that fetches and writes the file, or a reader of remote content: aws s3, gsutil, gcloud storage, az storage blob download, kubectl get -o/exec/cp, docker run/exec/cp, rclone, scp/rsync from a host, ssh host cmd, aws ssm/secretsmanager, vault kv get, ftp/lftp/tftp/smbclient, nc, a remote git ref; or PowerShell iwr/irm passed to iex, also inside powershell -c), over TLS from a named host is review and blocked for an agent; no TLS, a bare IP, a paste or short-link host, a decoder, or fetched code run through substitution, eval or a nested shell is deny. Reverse shells are deny. An obfuscated payload is review and blocked for an agent. A temp-dir executable alone (including ./x after cd into one) is review, not blocked. An unparseable command naming a fetch and a way to run code is review, blocked for an agent. A repository cloned and then built or run in the same command is recorded and allowed, because developer agents clone and build all day; it scores as a download-and-run only with no TLS, a bare public IP, or a paste or short-link host. The guard judges one command at a time. The armed kernel Execution Gate (Linux) can refuse unauthorized scripts and binaries.
ASI06
Memory & Context Poisoning
Limited / indirect
No detection of persistent-memory, RAG or context-store poisoning. Per-pod attribution can scope an investigation but does not contain the workload; this is indirect, not direct coverage.
ASI07
Insecure Inter-Agent Communication
Visibility only
Does not authenticate or verify agent-to-agent messages. Pro and Enterprise can attribute container events to pods and tenants, but it does not provide inter-agent isolation.
ASI08
Cascading Failures
Supporting
The circuit breaker and rate limits limit tool loops; Pro and Enterprise add an independent watchdog, containment and a kernel disarm kill-switch. These limit how far a fault propagates; they do not guarantee it stops.
ASI09
Human-Agent Trust Exploitation
Partial
Approval routing counts only when it carries independent evidence, a risk summary, explicit confirmation and an audit trail. A bare Telegram approval does not.
ASI10
Rogue Agents
Direct · kernel (Linux)
Reverse shells, cryptominers, system destruction and security-tooling tampering are deny. A scheduled task or service installed to run again is review and blocked for an agent, deny alongside other evidence; a write to a file that runs later (shell startup and logout files, ~/.ssh/rc, a PROMPT_COMMAND, git hooks, an autostart Exec= line, a systemd ExecStart= line, cron.d and crontab lines) is recorded, and the command it holds scores as the command it becomes: an export or alias line is allowed, a fetch-and-run or reverse shell is not. The armed kernel Execution Gate (Linux) can contain unauthorized execution.
Short answers
The ten risks, in one answer.
What are the OWASP Top 10 for Agentic Applications 2026 (ASI01 to ASI10)?
The ten risks the OWASP GenAI Security Project lists for agentic AI, in OWASP's order: ASI01 Agent Goal Hijack; ASI02 Tool Misuse & Exploitation; ASI03 Identity & Privilege Abuse; ASI04 Agentic Supply Chain Vulnerabilities; ASI05 Unexpected Code Execution; ASI06 Memory & Context Poisoning; ASI07 Insecure Inter-Agent Communication; ASI08 Cascading Failures; ASI09 Human-Agent Trust Exploitation; ASI10 Rogue Agents. The matrix on this page shows how InnerWarden covers each one, labelled direct, partial or supporting.
Scope by tier
What detects, what enforces, what we do not claim.
Free detects. Paid enforces.
The Community core (inner-warden, Apache-2.0) detects and advises on the runtime-observable risks; an agent that respects the verdict avoids most of them. Pro and Enterprise add kernel enforcement on Linux for tool misuse, unexpected code execution and rogue agents (ASI02 / ASI05 / ASI10). That enforcement sits below a jailbroken, non-cooperative agent, outside the prompt it obeys. Privilege controls support ASI03, and per-pod attribution gives ASI07 visibility, not isolation.
Where we're honest about scope
Some risks are architectural, not runtime, so a host guardrail only goes so far. ASI04 is look-alike package-name screening and runtime impact mitigation, not supply-chain validation (no SBOM or provenance). ASI06 is limited and indirect: we don't detect memory or RAG poisoning. ASI07 is visibility only, not isolation or message authentication. ASI09 counts only when the human sees independent evidence, not a bare approval. We label each row rather than claim more than the code does. A secret/PII redaction transform is an additional data-protection control, not an ASI claim.
The reason chain
A deny that names the risk it caught.
This is the layer the model cannot talk its way out of: on Linux the refusal is a kernel -EPERM, and a jailbroken agent cannot argue with it. And because every verdict maps to its risk class, a deny doesn't just say no. It says which agentic risk it caught. check-command returns the OWASP Agentic ids (["ASI05"] for a reverse shell) alongside the decision, so your security team sees it in the framework they already evaluate against.
Independent mapping to the OWASP Top 10 for Agentic Applications 2026. InnerWarden is not endorsed or certified by OWASP. ASI01-ASI10 and their titles are the property of the OWASP Foundation.