Back to skill

Security audit

volcengine-landing-zone

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because scanners flagged risks around targeting the intended cloud account and persistent privileges; the review is incomplete, so it does not establish that the skill is safe.

Consider the suspicious verdict and avoid installing until the account-targeting concern and undeclared capabilities have been reviewed. Scanner coverage was partial: 52 of 53 components were scanned, with one static analysis limit and unresolved path references, so the evidence is incomplete.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (10)

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · SKILL.md (reported line 80)May include surrounding context.

md
### G4. File Review Pause
**Trigger**: a phase produces a local file that the user needs to view, such as an initial password, login information, or a summary report.
**Action**: deliver the actual file using the Local File Display rule above, tell the user to review it, and pause there waiting for the next instruction.
**Forbidden**: do not skip the pause because you want to show multiple files later together. G2 does not exempt this pause. Do not echo the full initial password in chat, and do not replace file delivery with a text restatement.
**Self-check**: after producing a sensitive or user-review file, have I opened or delivered it according to the display protocol and then stopped? If not, stop.

Code capability 'env' detected in assets/blueprints/landing-zone-setup/04-log/tos_activate.py but not covered by declared permissions.

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'env' capability that is not covered by the declared 'allowed-tools' preapproval. Review the host runtime's actual enforcement policy; runtime permissions are unknown unless separately verified, and this declaration alone does not establish a permission ceiling.

Content

No source excerpt is available for this finding.

Code capability 'network' detected in assets/blueprints/landing-zone-setup/04-log/tos_activate.py but not covered by declared permissions.

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'network' capability that is not covered by the declared 'allowed-tools' preapproval. Review the host runtime's actual enforcement policy; runtime permissions are unknown unless separately verified, and this declaration alone does not establish a permission ceiling.

Content

No source excerpt is available for this finding.

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · references/landing-zone-setup/guidebook.md (reported line 74)May include surrounding context.

md
ts from default names or assume the fixed five-account shape.
- Record key phase outputs and summary-report material continuously in the background. Do not wait until all phases finish and reconstruct everything from memory.
- **Existing-resource reconciliation (all phases).** Any blueprint resource with a stable real-world identity (fixed name, service-linked role name, share name) may already exist outside this workspace's Terraform state: the customer deployed before, is redeploying, lost or reset state, a prior run partially applied, or the cloud auto-created it (for example, service-linked roles are often created on first use of NAT or Transit Router even in an otherwise untouched account). A blind create against an already-existing object fails with `AlreadyExists`, so this is a normal first-run concern, not only a recovery concern. Because the provider's data sources here are ID-only or cannot filter by name, discovery is a read-only `ve`/OpenAPI step, not a Terraform data looku

The comment guarantees that an incorrect account_id fails, but the provider simply assumes a role in whichever account_id is supplied.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

AssumeRole succeeds for any supplied account ID whose named role trusts the base identity; it does not establish that the account is the intended deployment target. In a multi-account organization, a mistaken but accessible ID can therefore direct network changes into another account, contrary to the comment's guarantee. Phase-scoped confirmation reduces this risk but does not make the provider fail closed.

Content

No source excerpt is available for this finding.

The comments promise that an incorrect account_id cannot target the wrong account, but the provider assumes whichever accessible account the input identifies.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The template selects the destination directly from account_id without independently establishing that it matches the intended member account. A different accessible member account can accept the same role assumption, allowing baseline resources to be deployed there despite the comment's assurance. As a reusable template, this misleading guarantee can propagate into additional baselines.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/landing-zone-setup/01-organization.md (reported line 36)May include surrounding context.

md
## Minimum Input

- Prefer read-only lookup for `root_ou_id`; do not ask the user for it first.
- Ask for `prefix` only when it is missing. The default architecture uses it for account names, and later phases reuse it for resource naming.
- Ask for `region` only when it is missing. Prefer `VOLCENGINE_REGION`, otherwise default to `cn-beijing`.
- Resolve and persist the verified management account ID. Persist the organization ID when discovery exposes it.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/landing-zone-setup/02-finance.md (reported line 43)May include surrounding context.

md
## Result Notes

- After this phase completes, make it clear which core accounts already have the intended finance relationships
- Finance relationships are not automatically removed when Terraform resources are destroyed. If removal is needed, run `ve billing DeleteFinancialRelation` or an equivalent cleanup command separately

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/landing-zone-setup/guidebook.md (reported line 30)May include surrounding context.

md
- Match an OU by exact parent path plus exact name and an account by explicit ID or unique exact `account_name`. Populate generic `existing_ou_ids` and `existing_account_ids` with unambiguous stable-key matches.
- Classify each desired object as `reuse`, `create`, or `conflict`. A duplicate match, wrong parent, account name/display mismatch, or existing account in the wrong OU is a conflict. Do not silently move an account or choose one duplicate.
- Inspect Terraform state before injecting reuse IDs. A resource already managed by the workspace remains managed and must not also be converted into an unmanaged data reference.
- Ask the user for an ID only when read-only discovery cannot resolve an ambiguity. Do not ask the user to author the architecture object or internal stable keys.

### Step 1: Prepare the Workspace Execution Copy

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/preflight-checks.md (reported line 107)May include surrounding context.

md
- `Solution Deployment`
  - Confirm the package source is resolvable and materialize it under `${WORKSPACE_ROOT}/solution-deployment/packages/<name>/`.
  - Resolve the package entrypoints, required tools, declared targets, inputs, and writes.
  - For every Terraform target, resolve either provider role assumption or a runtime credential/profile that already represents the target. Do not ask the user to choose when the package and runtime context make the strategy clear.
  - For a verified member account in the current enterprise organization, when cross-account role assumption is required and no role is specified, use `trn:iam::<account-id>:role/OrganizationAccessControlRole`. Do not infer this role for an account outside the current organization.
  - Verify read-only that each effective Terraform identity resolves to its declared target before planning any write. Stop on an unverifiable or mismatched target-to-identity mapping.
  - Confirm execution will happen under `solution-deploym