Lodestar and the OWASP Top 10 for Agentic Applications¶
The OWASP GenAI Security Project published the OWASP Top 10 for Agentic Applications for 2026 in December 2025 (ASI01–ASI10). This page maps each category to what Lodestar does about it, and says plainly where it does little or nothing.
Category names below are taken from the official PDF (CC BY-SA 4.0), not from secondary summaries.
How to read the ratings¶
Lodestar is a governance layer around an agent, not an agent runtime and not a model-level defence. It constrains what a compromised or confused agent can believe and do; it does not stop the model from being manipulated.
- Addressed — a Lodestar component enforces a control for the category's core scenario, and at least one named probe pins it.
- Partial — a named probe pins part of the category; the rest is stated as a gap.
- Not addressed — no probe pins it. Adjacent signals are listed where they exist, but they are not claimed as coverage.
Probes are executable specs, so every claim here can be checked by running the
probe. They live in packs/lodestar-core/ and packs/coding-agent-safety/
(run with lodestar harness run --pack <name>). Several guarantees are
TypeScript-level boundaries, not OS- or network-level containment; the
per-adapter ADRs in
.claude/adr/ say
which.
Summary¶
| ID | Category (OWASP wording) | Rating |
|---|---|---|
| ASI01 | Agent Goal Hijack | Partial |
| ASI02 | Tool Misuse and Exploitation | Addressed |
| ASI03 | Identity and Privilege Abuse | Partial |
| ASI04 | Agentic Supply Chain Vulnerabilities | Partial |
| ASI05 | Unexpected Code Execution (RCE) | Partial |
| ASI06 | Memory & Context Poisoning | Addressed |
| ASI07 | Insecure Inter-Agent Communication | Not addressed |
| ASI08 | Cascading Failures | Partial |
| ASI09 | Human-Agent Trust Exploitation | Partial |
| ASI10 | Rogue Agents | Not addressed |
ASI01 — Agent Goal Hijack · Partial¶
Risk. Attackers redirect an agent's objectives, task selection, or decision pathways, typically through instructions hidden in content the agent reads.
What Lodestar does. It does not try to detect the hijack inside the model.
It limits what injected content can become: content from documents, web
pages, and tool results is recorded as external_document evidence, and the
auto-observation gate refuses to let that evidence
promote a claim to supported. Injected "facts" therefore stay unverified
and out of trusted context. Actions the hijacked agent then proposes still
meet the trust ladder and its approval holds.
Probes. prompt-injection-cross-tool and
poisoned-file-cannot-hijack-feature-work (coding-agent-safety);
mcp-proxy-injection-defense and auto-observation-gate (lodestar-core).
Gap. The agent can still be talked into proposing a harmful action; the defence is that the proposal is gated, not that it is never made. Lodestar has no model-side goal-integrity check.
ASI02 — Tool Misuse and Exploitation · Addressed¶
Risk. Agents misuse legitimate tools through prompt injection, misalignment, or unsafe delegation, leading to exfiltration or workflow hijacking.
What Lodestar does. The Action Kernel runs every tool call in two phases (propose, then execute) against a typed action contract with an explicit trust level. The Policy Kernel gate is three-valued (allow / deny / hold) and defaults to deny. Each native adapter (shell, git, http, messaging, sql, filesystem, nostr, payments, vector) pins its own invariants: argv-only execution, operator-pinned destinations, bounded output, scoped credentials. See governed egress.
Probes. l4-action-requires-approval, pending-approval-cannot-execute,
unmatched-action-defaults-to-deny, granted-approval-still-revalidates-preconditions
(gate and kernel); shell-adapter-enforces-sandbox-invariants,
http-adapter-enforces-egress-invariants,
messaging-adapter-enforces-egress-invariants,
sql-adapter-enforces-invariants,
filesystem-adapter-enforces-write-invariants,
payment-adapter-enforces-send-invariants (adapters).
Gap. Coverage is for tools that go through the kernel or the MCP proxy. A tool the agent calls natively, outside Lodestar, is invisible to it.
ASI03 — Identity and Privilege Abuse · Partial¶
Risk. Exploiting delegation chains, inherited credentials, and cached context to escalate access.
What Lodestar does. The trust-ladder floor means a policy rule cannot
grant an action less scrutiny than its rung requires. Approvals are signed
(Ed25519) and verified against operator-pinned approver keys, so a forged or
tampered grant cannot release a held action. Signature verification applies
wherever approver keys are pinned; an operator can explicitly opt out with
allow_unsigned for trusted local setups. Where an operator configures a
policy rule with quorum >= 2, an M-of-N quorum is adjudicated by the kernel
over each approver's own signed resolution, counting only eligible, distinct
approvers and excluding the action's own proposer. Quorum is opt-in per rule,
not automatic for high-stakes actions. Adapter credentials
are operator-supplied, scoped, and kept out of inputs, observations, and
errors.
Probes. ladder-floor-overrides-allow-rule, forged-approval-cannot-execute,
policy-version-signature-required,
quorum-requires-distinct-eligible-approvers.
Gap. Lodestar governs the approval identity and the credentials its own adapters hold. It does not manage the agent's identity, credential lifecycle, or delegation chains across systems.
ASI04 — Agentic Supply Chain Vulnerabilities · Partial¶
Risk. Third-party tools, plug-ins, models, MCP servers, agents, and registries that are malicious, compromised, or tampered with.
What Lodestar does. This is covered for probe packs, Lodestar's own extension artifact. Pack manifests are Ed25519-signed with a per-file content digest, verified against consumer-pinned author keys before any pack code runs. Sources must pin an immutable version (npm integrity, full git SHA); resolution never runs install scripts or git hooks. Discovery indexes and verification badges are advisory and never confer trust. See the registry threat model and ADR-0016 through ADR-0021.
Probes. pack-manifest-signature-required, forged-pack-cannot-load,
tampered-pack-content-cannot-load, mutable-git-ref-rejected,
resolution-runs-no-pack-code, pack-index-signature-required,
unverified-badge-not-trusted.
Gap. Lodestar does not vet the agent's own dependencies, model weights, or downstream MCP servers. For MCP, it governs what a server's tools can do and what their output can become, not whether the server is trustworthy.
ASI05 — Unexpected Code Execution (RCE) · Partial¶
Risk. Agent-generated or tool-triggered code escalating to remote or local code execution.
What Lodestar does. The shell adapter enforces an allowlist, argv-only
execution (no shell interpolation), a scoped environment with no host-env
passthrough, a wall-clock timeout, and bounded output. Probe packs run under
a scoped environment and, for external packs, inside an OS sandbox
(sandbox-exec on macOS, bubblewrap on Linux).
Probes. shell-adapter-enforces-sandbox-invariants,
runner-denies-host-env-to-probe,
runner-sandboxes-probe-filesystem-and-network.
Gap. The shell adapter's sandbox is a TypeScript-level boundary, not an OS sandbox: it governs how commands are constructed and what environment they get, not what an allowed binary can do once running. There is no sandbox for arbitrary agent-generated code.
ASI06 — Memory & Context Poisoning · Addressed¶
Risk. Adversaries corrupt or seed the information an agent retains, retrieves, or reuses (summaries, embeddings, RAG stores), biasing later reasoning.
What Lodestar does. This is the category the Memory Firewall
exists for. Memory state is split across four orthogonal axes (truth,
retrieval, security, freshness); agent-written memories cannot self-promote;
external_document and model_inference evidence cannot move a claim to
supported on its own; quarantined beliefs are not retrievable;
sensitivity ceilings gate what leaves a session. Poisoned content stays
inert across sessions, including through the Postgres-backed stores, and RAG
chunks are treated as untrusted evidence as well. See the
memory-poisoning threat model and
reading isn't believing.
Probes. memory-poisoning-basic, auto-observation-gate,
quarantined-not-retrievable, guard-import-no-self-promote,
reflection-cannot-promote-to-normal-alone,
vector-retrieval-cannot-auto-promote, world-model-withholds-gated-current-state,
sensitivity-ceiling (lodestar-core); tool-poisoning-cross-session
(coding-agent-safety; needs a Postgres test database, skips loudly without
one).
Gap. Coverage applies to memory and context that pass through the firewall. Context the host keeps outside it (a raw chat transcript, a vendor memory feature) is not governed.
ASI07 — Insecure Inter-Agent Communication · Not addressed¶
Risk. Weak authentication, integrity, or authorization between agents lets attackers spoof, intercept, or manipulate agent-to-agent messages.
What Lodestar does. Nothing specific. Lodestar has no agent-to-agent messaging layer, and no probe pins one. Content that arrives from another agent through a governed tool is treated like any other tool output (untrusted evidence), which is incidental protection, not a control for this category.
ASI08 — Cascading Failures · Partial¶
Risk. A single fault (hallucination, malicious input, corrupted tool, poisoned memory) propagates across agents and compounds into system-wide harm.
What Lodestar does. It limits propagation of epistemic faults along the
chain. When a belief is contradicted, the reflection pass proposes flagging
the decisions that depended on it. Separately, sentinel alerts (for example on
a suspicious memory origin) can hold dependent actions at pending_approval
instead of letting them execute.
Probes. contradicted-belief-flags-dependent-decisions,
sentinel-alert-gates-dependent-action, guard-arbiter-gates-dependent-action,
mcp-proxy-arbiter-gates-dependent-action.
Gap. There is no circuit breaker, rate limiting, or blast-radius cap across agents or workflows, and no probe for multi-agent failure propagation.
ASI09 — Human-Agent Trust Exploitation · Partial¶
Risk. Agents exploit human trust (authority, fluent explanations) to get risky actions approved, or to mask what they are doing.
What Lodestar does. Approval is an explicit, out-of-band, signed event, so a forged or tampered approval cannot release a held action when approver keys are pinned (the unsigned opt-out is explicit), and a held action that times out is denied. Separating approver from proposer is enforced by the kernel only for quorum-enabled holds; on the single-approver path it depends on the resolver being a separately authenticated channel, not the agent. The decision chain is recorded and can be reported on afterwards.
Probes. forged-approval-cannot-execute,
forged-approval-via-http-channel-cannot-execute, approval-timeout-denies,
pending-approval-cannot-execute.
Gap. Signed approval authenticates who approved, not whether they understood what they approved. The approver is not guaranteed to see the tool and inputs of the action in front of them, so a persuasive agent can still shape a rubber-stamp. This is a known gap we intend to close, and no probe pins it yet. Opt-in quorum raises the cost of such an attack but does not remove it.
ASI10 — Rogue Agents · Not addressed¶
Risk. Agents that drift or become misaligned and act outside their intended scope without an active attacker.
What Lodestar does. Nothing that contains a rogue agent. Action
containment (ASI02) bounds what any agent can do, whether or not it is
rogue, but that is not detection or response. The closest detective signals
are calibration reports and the confidence-drift probe; they surface
over-confidence in an agent's beliefs and are not a rogue-agent detector.
Maintenance¶
Ratings are claims about the code at a given release. If a probe named here is renamed or removed, or a gap is closed, update this page in the same change.