# OWASP Agentic Top 10 — coverage

> How Open Agent Primitives maps against the OWASP Top 10 for Agentic Applications (2026) — what each primitive addresses, and where the gaps are.

Source: https://openap.org/docs/owasp-top10

A map of how OAP addresses the **OWASP Top 10 for Agentic Applications (2026)**. Each concept
in these docs is designed around specific threats below; each threat here names the OAP mechanism that
answers it and the gaps that remain. This is an *advisory* assessment, not a certification — coverage
is a qualitative judgement, and honest about what's missing.

Across the ten: **4 substantially addressed · 6 partial.**

> **Note: How to read this**
>
> Each item links back to the concept that addresses it, and each concept's overview links here. "Substantial"
> means the threat is a design center of OAP; "Partial" means real defenses with named gaps.

### ASI01 — Agent Goal Hijack (coverage: Partial)

**What OAP does.** Tool output cannot widen what the agent may do: every resulting action is checked outside
the model, and high-impact and *external* actions can be human-approval-gated. Tool results are the *only*
non-user ingress path; as a secondary aid, each is wrapped in an unpredictable per-result nonce and the system
prompt's spotlighting rule tells the model to treat delimited content as data, never instructions. A pluggable
content-guard seam inspects tool I/O, and two inspectors ship: a prompt-injection classifier (run in a
zero-egress pod) and a URL allowlist. — see [Safe tools](https://openap.org/docs/safe-tools),
[Content guards & egress](https://openap.org/docs/content-guards-egress), and [Agent definition](https://openap.org/docs/agent-definition).

**Gaps.** No goal-lock or plan-divergence detection; inspection covers tool I/O, not the agent's own replies.

### ASI02 — Tool Misuse & Exploitation (coverage: Substantial)

**What OAP does.** This is OAP's home turf. Tools are dual-validated — at authoring time and again at
execution time — behind deny-by-default subcommand/field allowlists and CEL, and they run as **argv arrays,
never through a shell**. Each tool's `stateImpact` (readonly / readwrite / external) drives the authorization
check and whether a human is asked; every action is checked per-action against SpiceDB. Secrets are scrubbed
from output; default-on circuit breakers, plus opt-in per-tool rate limits and per-call data-volume budgets,
bound runaway use. — see
[Safe tools](https://openap.org/docs/safe-tools).

**Gaps.** Egress control is L3/L4 only; the volume budget is per-call, not cumulative.

### ASI03 — Identity & Privilege Abuse (coverage: Substantial)

**What OAP does.** Each agent has its own `AgentIdentity` — there is no shared credential catalog. Acting on a
user's behalf goes through `UserIdentity → SessionUserIdentity` with the credential set narrowed to what the
session's agent asked for, and a token broker resolves credentials just-in-time (a valet key, fail-closed).
Critically, **an agent never acts on authority of its own**: checks run against the user it acts for, or, for a
session with no user, a principal of its own (the session itself, or a service it declares) with no standing grants — which structurally blocks the
confused-deputy problem; grants are bound to an arguments-hash HMAC caveat, and credential revocation reaches in-flight
sessions. — see [Identity](https://openap.org/docs/identity).

**Gaps.** Sidecar credentials are frozen at pod-create (no mid-execution re-auth); revocation delivery is
at-most-once.

### ASI04 — Agentic Supply Chain (coverage: Partial)

**What OAP does.** The MCP tool manifest is content-hash-pinned, with pluggable pinning kinds (mcp / image /
cli / skill) under a tiered policy, and drift is *enforced*, not merely recorded. A tiered `allowedMCPServers`
allowlist bounds what an agent may install. — see [Safe tools](https://openap.org/docs/safe-tools).

**Gaps.** No descriptor signing or attestation, and no SBOM/AIBOM — a hash pins *content*, not *origin*.

### ASI05 — Unexpected Code Execution (RCE) (coverage: Substantial)

**What OAP does.** Tool pods are hardened: non-root, read-only root filesystem, all capabilities dropped,
seccomp `RuntimeDefault`, no service-account token. Tools execute as **argv arrays over pod exec — never a
shell, never `eval`** — against a closed set of tools, in per-session sandbox pods. — see
[Safe tools](https://openap.org/docs/safe-tools).

**Gaps.** Egress is L3/L4 only. Isolation is per session: tool calls within one session share a UID, /proc,
/tmp and /work.

### ASI06 — Memory & Context Poisoning (coverage: Partial)

**What OAP does.** Memory is authorized per-subject through SpiceDB — read is session membership, write and
delete are the creator's — with tenant/scope isolation and search post-filtered through SpiceDB. External
entries are tagged untrusted, and the transcript, authz decisions, and audit kinds are an **append-only,
Ed25519-signed, tamper-evident log** verifiable offline with `oap audit verify`. — see [Memory](https://openap.org/docs/memory) and
[Authorization](https://openap.org/docs/authorization).

**Gaps.** No content validation on memory writes. Assistant turns are ingested into the knowledge graph without
separate validation, so a poisoned turn can influence later recall.

### ASI07 — Insecure Inter-Agent Communication (coverage: Partial)

**What OAP does.** Delegation reaches only a **closed, reviewed subagent roster**, validated as a DAG, with
entries that can be pinned to a bundle digest. The delegation tree is bounded (`maxDelegatedAgents`, and a
per-member mode ceiling), and each subagent is its own scoped, budgeted, audited AgentSession. Parent and child
exchange only three typed, separately authorized directions — `ask_parent`, `return_result`, and
`reply_to_subagent` — delivered as inspected tool results; free-form agent-to-agent messages were deliberately
removed. The sender is authenticated by its own per-session bus subject, and the destination must match a real
Channel joining the two sessions. Passing data past its audience parks for the data owner's approval. — see
[Subagents & delegation](https://openap.org/docs/subagents-delegation) and [Channels](https://openap.org/docs/channels).

**Gaps.** Delegation stays within one cluster; there is no protocol for agents outside it.

### ASI08 — Cascading Failures (coverage: Partial)

**What OAP does.** Per-session isolation contains blast radius; a silence watchdog, per-phase budgets, and
default-on circuit breakers (a per-tool/per-origin three-state machine) stop runaway loops; default-on egress
NetworkPolicies bound each session and sandbox; the operator's Secret/ConfigMap reads are code-confined to
adopted objects with a fail-loud tripwire. — see [Agent definition](https://openap.org/docs/agent-definition).

**Gaps.** The breaker is per-tool/origin, not cross-cutting; the operator's `secrets` RBAC is cluster-wide by
default.

### ASI09 — Human-Agent Trust Exploitation (coverage: Substantial)

**What OAP does.** OAP's whole approval design targets this. External/high-impact actions require explicit
confirmation, and the card a human reads is built by an **injection-isolated summarizer** — a second LLM that
sees only the operator-signed schema and arguments, never the primary chat. The mechanically-derived "What" is
shown separately from the agent's "Why", a permanent Show-Details escape hatch exposes the raw arguments, and
previews are separated from effects (an SVG artifact renders as an inert `data:` URI, no script). The approver
is session-derived and owner-gated at click time. — see [Channels](https://openap.org/docs/channels) and the
[Authorization](https://openap.org/docs/authorization) guides.

**Gaps.** No plan-divergence cross-check; no confidence-weighted UI cues.

### ASI10 — Rogue Agents (coverage: Partial)

**What OAP does.** Sessions are disposable and scoped, checked per-action against SpiceDB; the declared tool
set is an **enforced behavioral manifest** — an undeclared tool is structurally uncallable. The audit system
*is* the audit trail: every authorization check writes a tamper-evident, signed decision record that outlives
the session. Runtime scope can be narrowed, and both credentials and tools can be revoked into in-flight
sessions. — see [Authorization](https://openap.org/docs/authorization) and [Memory](https://openap.org/docs/memory).

**Gaps.** The tool manifest isn't signed/attested and drift within the allowed surface is undetected;
tamper-*evidence* is not tamper-*proofing*; there is no self-replication guard.

## Mapped to the primitives

Each concept is designed around specific threats:

- **[Safe tools](https://openap.org/docs/safe-tools)** → [ASI02](https://openap.org/docs/owasp-top10#asi02), [ASI04](https://openap.org/docs/owasp-top10#asi04), [ASI05](https://openap.org/docs/owasp-top10#asi05), [ASI01](https://openap.org/docs/owasp-top10#asi01)
- **[Identity & credentials](https://openap.org/docs/identity)** → [ASI03](https://openap.org/docs/owasp-top10#asi03)
- **[Authorization](https://openap.org/docs/authorization)** → [ASI02](https://openap.org/docs/owasp-top10#asi02), [ASI03](https://openap.org/docs/owasp-top10#asi03), [ASI06](https://openap.org/docs/owasp-top10#asi06), [ASI10](https://openap.org/docs/owasp-top10#asi10)
- **[Agent definition](https://openap.org/docs/agent-definition)** → [ASI01](https://openap.org/docs/owasp-top10#asi01), [ASI07](https://openap.org/docs/owasp-top10#asi07), [ASI08](https://openap.org/docs/owasp-top10#asi08), [ASI09](https://openap.org/docs/owasp-top10#asi09)
- **[Channels & continuity](https://openap.org/docs/channels)** → [ASI07](https://openap.org/docs/owasp-top10#asi07), [ASI09](https://openap.org/docs/owasp-top10#asi09)
- **[Memory & knowledge](https://openap.org/docs/memory)** → [ASI06](https://openap.org/docs/owasp-top10#asi06), [ASI10](https://openap.org/docs/owasp-top10#asi10)

---

This page is based on the [OWASP Top 10 for Agentic Applications (2026)](https://genai.owasp.org)
by the OWASP GenAI Security Project, licensed under
[CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/). The assessment is AuthZed's own and is not
affiliated with or endorsed by OWASP.
