Back to skill

Security audit

Rent A Logged In Agent

Security checks for vulnerabilities and agentic risk

Overview

This skill openly enables renting out a logged-in coding agent to remote callers, but that design gives other people prompt-level access to a credential-bearing agent and depends on an unpinned external CLI.

Install only if you fully trust SettleMesh and every allowed caller, and run it from a dedicated, disposable account or VM with no unrelated SSH, cloud, repository, or personal credentials. Avoid the no-sandbox override, pin and verify the CLI version if possible, and treat any allowed remote caller as capable of spending your agent quota or trying to extract credentials from the lent session.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (2)

T08 · Insecure Dependencies

Error
Location
SKILL.md:28
Finding

Unpinned Global Installation of an External npm Package

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 28-31
Vulnerability Type: Unpinned third-party dependency installed globally
Risk Level: High

Vulnerable Code Snippet

markdown
## Setup (once)

1. `npm install -g settlemesh@latest`
2. `settlemesh login` (a human approves once in the browser).
3. Have `claude-code` or `codex` installed **and logged in** on this machine.

Technical Analysis

The setup instructions install settlemesh@latest globally without an exact version, integrity hash, lockfile, or locally reviewable implementation. The effective code can therefore change after this Skill has been reviewed. Global npm installation may also execute package lifecycle scripts with the privileges of the user performing the installation.

The project contains only SKILL.md, so the external CLI's authentication, credential handling, sandbox enforcement, authorization, network communication, and billing behavior cannot be independently verified from the audited artifact.

This is an insecure supply-chain practice rather than proof that the current SettleMesh package is malicious. Nevertheless, it creates a direct execution path from a mutable third-party package registry to the user's machine.

Attack Path

  1. An attacker compromises the npm publisher account, package distribution infrastructure, or a future release of the dependency.
  2. The compromised package is published under the latest tag.
  3. A user follows the Skill's setup instructions and runs npm install -g settlemesh@latest.
  4. npm downloads the attacker-controlled release and may execute its lifecycle scripts.
  5. The installed CLI subsequently executes with the user's permissions and can access data and resources available to that user.

Impact Assessment

Successful exploitation could permit arbitrary command execution under the installing user's account. Depending on that account's permissions and environmen ...[truncated 324 chars]

Remediation
View remediation

Remediation Suggestions

  • Pin the dependency to an exact, reviewed version rather than using @latest.
  • Publish and verify the expected package integrity hash.
  • Use a project-local installation with a committed lockfile instead of a global installation.
  • Disable npm lifecycle scripts during installation where operationally possible, then explicitly run only reviewed setup steps.
  • Document the authoritative package publisher, registry, expected signatures, and release verification procedure.
  • Include or link to auditable source corresponding exactly to the pinned release.
  • Run the CLI in a dedicated, minimally privileged account or isolated environment without access to unrelated host credentials.
  • Require a security review before upgrading the pinned dependency.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:61
Finding

Remote Prompt Execution with Readable Agent Credentials and an Unsafe Sandbox Bypass

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 61-72
Vulnerability Type: Excessive credential exposure and optional removal of filesystem confinement
Risk Level: Critical

Vulnerable Code Snippet

markdown
## Security — read this before lending

- **Throwaway sandbox HOME.** Each job runs in a fresh sandbox HOME with your
  login injected; your host `~/.claude` keychain / `~/.codex` is **never mounted**.
- **Why the offer is never public.** The injected credential is readable by the
  prompt-driven agent, so the offer is gated to your allowlist / friend graph and
  **does not appear in discovery**. Never lend to strangers.
- **FAIL-CLOSED.** SettleMesh **refuses to lend** on a host with no
  filesystem-confining sandbox backend — it needs `sandbox-exec` (macOS) or
  `bwrap` (Linux). No sandbox, no lend.
- **`--i-accept-no-sandbox` is DANGEROUS** — a last resort only. It overrides the
  refusal and lets the lent agent read `~/.ssh`, `~/.aws`, and its injected login
  and leak them to the caller. Prefer installing `bwrap` (Linux, needs user
  namespaces) over using this flag.

Technical Analysis

The Skill's core design allows a remote caller to supply a prompt to a local coding agent while an injected login is readable by that prompt-driven agent. An allowlist or friend graph limits who may submit jobs but does not establish a security boundary against a malicious authorized caller, a compromised authorized account, prompt injection, or credential-extraction behavior.

The documented --i-accept-no-sandbox option is especially dangerous because it deliberately disables the fail-closed filesystem-confinement requirement. The Skill explicitly states that this mode permits the remote-prompt-driven agent to read host credential locations such as ~/.ssh and ~/.aws, as well as the injected agent login, and potentially return their contents to the caller.

This exceeds minimum privi ...[truncated 2578 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove or technically disable --i-accept-no-sandbox for credential-bearing lending sessions.
  • Never place reusable credentials in storage readable by the prompt-driven agent.
  • Replace reusable logins with short-lived, single-job, least-privileged capability tokens.
  • Mediate authenticated operations through a narrowly scoped broker so the agent can request approved actions without reading raw credentials.
  • Restrict every job to an ephemeral filesystem containing only explicitly selected input files.
  • Deny access to the host home directory, SSH agents, cloud metadata services, credential stores, Unix sockets, and unrelated environment variables.
  • Apply outbound network allowlists and response filtering to reduce credential-exfiltration channels.
  • Use a dedicated unprivileged operating-system account or disposable virtual machine for the worker.
  • Require explicit human confirmation before each lending session and for every expansion of filesystem, network, or credential access.
  • Treat friend and allowlist accounts as potentially compromised; implement per-job authorization, rate limits, audit logs, revocation, and anomaly detection.
  • Independently verify the external CLI's fail-closed behavior and sandbox policy before relying on its security claims.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (3)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The manifest description says to use this skill when the user wants to "share compute for pay" or "let a partner or friend invoke" their agent, which are broad natural-language triggers rather than a narrow invocation condition. It does not provide explicit trigger phrases, constraints, or negative examples to distinguish when this skill should activate versus more general sharing, compute, or collaboration requests.

Content

No source excerpt is available for this finding.

Unrestricted Tool Access

Medium
Category
Excessive Agency
Confidence
70% confidence
Finding

Skill grants unrestricted tool access without appropriate constraints. An agent with unfettered tool access can perform arbitrary actions including file modification, network requests, and code execution.

Content

Scanner excerpt · SKILL.md (reported line 48)May include surrounding context.

md
## The caller side (invoke)

The offer's address is the **lender's SettleMesh login email** — allowlist offers
are private and never show up in `services search` / `worker-offers list`, so the
caller invokes it directly:

Session Persistence

Medium
Category
Rogue Agent
Confidence
92% confidence
Finding

The skill is explicitly designed to let remote callers run prompts through a locally logged-in coding agent, and the document states that a live login is injected into each sandboxed job. Even with sandboxing, this creates a delegated-session model where an untrusted or insufficiently trusted caller can induce actions through the lender's authenticated agent, potentially causing data exfiltration, account misuse, or spend under the lender's identity.

Content

Scanner excerpt · SKILL.md (reported line 53)May include surrounding context.

caller invokes it directly:

bash
settlemesh worker invoke calleelqy@gmail.com --input '{"prompt":"Write a Python is_prime(n). Only code."}'

The caller supplies only {"prompt":"..."}. They find your address with