Back to skill

Security audit

Community Onboarding

Security checks for vulnerabilities and agentic risk

Overview

The skill mostly matches its course-creation purpose, but its update flow asks the agent to run server-provided commands that can modify the CLI or installed skill.

Review any install or update command before approving it. Only proceed if it is clearly a PersonWise CLI or skill update from the expected PersonWise origin, and be aware that selected documents/images may be uploaded to PersonWise to create the course.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:113
Finding

Server-Controlled Update Command Execution

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 113–128
Vulnerability Type: Command injection through trusted remote update responses
Risk Level: High

Vulnerable Code

text
- If `updates.cli.status` or `updates.skill.status` is `update_available` or `below_minimum`,
  tell the user once which component is outdated (installed versus latest) and quote the exact
  `action` command. The task cannot continue until that update is installed. Ask for approval;
  with approval, run exactly that command (it already carries the required `--approve-upgrade`
  argument).
...
When the printed `action` refreshes this installed Skill, replace `<skill-directory>` with the
directory of this installed Skill (the directory containing this Skill's SKILL.md). Never run
`doctor` or a generic capability preflight to check freshness; the `update check` command above is
the freshness check. Never ask more than once per component per session, and never substitute
another command, flag, origin, or download path for the printed `action`.

Technical Analysis

The Skill instructs the agent to execute an action command supplied by a remote CLI or service response without independently validating its command structure. There is no local allowlist of permitted executables and subcommands, no structured argument validation, and no rejection of shell operators or unexpected paths.

Although user approval is required, that approval only confirms the general update operation. It does not ensure that the remotely returned command is limited to installing an authentic PersonWise update. The instruction to run the command “exactly” and not substitute it makes the remote response an effective command-execution control channel.

Attack Path

  1. An attacker compromises or impersonates a component capable of influencing the PersonWise update response, such as the service, CLI response-processing path, or authenticated ...[truncated 1366 chars]
Remediation
View remediation

Remediation Suggestions

  • Do not execute command strings received from a remote service or parsed CLI response.
  • Change the update response to structured metadata, such as component, target version, artifact digest, and operation identifier.
  • Map that metadata locally to a strict allowlist of fixed commands shipped with the Skill.
  • Invoke the approved executable directly with an argument array rather than through a shell.
  • Permit only expected update operations and reject shell metacharacters, unknown flags, unexpected paths, and unrecognized origins.
  • Require the destination to resolve to the expected CLI binary or installed Skill directory and reject symlinks or path traversal.
  • Pin the release origin and verify signatures or exact digests before installation.
  • Present the validated component, version, source, and destination to the user before requesting approval.
  • Keep the bundled bootstrap scripts as the only CLI installation path where possible, because they use fixed origins and pinned hashes.
  • If remote update metadata fails validation, stop the workflow and report the failure instead of falling back to executing the supplied action.
Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (4)

Scope Creep

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The script performs an outbound network download via curl even though the skill metadata declares only Bash permission. This creates a permission/scope mismatch: invoking the skill can fetch and install remote code, expanding capability beyond what a reviewer or user may reasonably expect from the declared permissions. The script does include TLS, size limits, and checksum verification, which reduces supply-chain risk, but it does not eliminate the policy and trust-boundary issue.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill hard-codes use of an international PersonWise service by default and forbids switching based on other signals unless the user explicitly asks. Defaulting to a specific market/service without explicit opt-in can route authentication and course data to a jurisdiction or service context the user did not intend, creating privacy, compliance, and data residency risks.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill explicitly instructs the agent to run install and upgrade commands, while also telling it not to provide a separate user-facing warning on the normal path. Even though the metadata mentions explicit approval and the host may enforce approval, the workflow normalizes code installation/update from bundled scripts, which increases the chance of silent environment modification or supply-chain risk if approvals are unclear or inconsistently enforced.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
80% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · assets/bootstrap.sh (reported line 165)May include surrounding context.

sh
echo "Release executable checksum mismatch." >&2
  exit 6
fi
chmod 700 "$candidate"
if [ "$(uname -s)" = "Darwin" ] && [ "$native_signature_status" = "verified" ]; then
  codesign --verify --deep --strict --verbose=2 "$candidate"
  spctl --assess --type execute --verbose=2 "$candidate"