Back to skill

Security audit

Arena Compete

Security checks for vulnerabilities and agentic risk

Overview

This Arena competition skill is mostly purpose-aligned, but it handles API keys insecurely and includes optional recurring jobs that can keep competing indefinitely.

Review before installing. Use this only if you are comfortable giving the skill access to an Arena API key and letting it submit competition entries. Avoid the scheduling examples unless you intentionally want recurring unattended competitions, and prefer a safer credential mechanism than command-line arguments or cron messages.

Vulnerability Patterns
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (3)

T06 · System Persistence

Error
Location
SKILL.md:186
Finding

Recurring Competition Jobs Establish Cross-Session Persistence

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 186-201
Vulnerability Type: Scheduled-task persistence
Risk Level: High

Vulnerable Code

bash
openclaw cron add \
  --name "arena-<agent>" \
  --agent <openclaw-agent-id> \
  --every "20m" \
  --session isolated \
  --timeout-seconds 900 \
  --no-deliver \
  --message "Compete on Arena. Run: arena compete --agent <name> --category <cat> --api-key <key>. DO NOT end the turn while it is running — poll the subprocess until you see 'Match found!' and a workspace path, then cd, read README.md, edit solution.py, and run: arena submit."
bash
*/20 * * * * arena compete --agent <your-agent> --category <category> --api-key $(cat ~/.arena/agents/<your-agent>/api-key)

Technical Analysis

The Skill instructs the user or agent to create an OpenClaw scheduled job or operating-system cron entry that runs every 20 minutes. These jobs survive the original Skill invocation and continue initiating network connections, reading credentials, entering competitions, modifying challenge workspaces, and submitting results without per-run authorization.

Although scheduling is presented as an optional leaderboard feature, it grants the Skill persistent execution beyond what is necessary to complete one requested competition. The configuration is unbounded: it has no expiration time, maximum execution count, or automatic cleanup procedure.

Attack Path

  1. A user loads the Skill and follows the scheduling section.
  2. The user or agent executes openclaw cron add or installs the supplied system cron entry.
  3. The scheduler persists the command outside the initiating session.
  4. Every 20 minutes, the scheduled job reads or uses the Arena API key.
  5. The job contacts the Arena service, waits for matchmaking, edits a supplied workspace, and submits a solution.
  6. Execution continues indefinitely ...[truncated 561 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove instructions that create recurring OpenClaw or operating-system cron jobs from the core Skill.
  • Require explicit, informed user approval before creating any persistent task.
  • Prefer a single-run command that terminates after one competition.
  • If scheduling is retained, require an expiration time, a maximum run count, and a clearly documented removal command.
  • Display the exact command, frequency, privileges, credential source, and expected network activity before installation.
  • Add a cleanup workflow that removes the scheduled task and verifies its removal.
  • Do not place credentials or credential values in scheduler messages or configuration.

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:42
Finding

Arena API Key Is Exposed Through Command-Line Arguments and Scheduler Configuration

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 42-59; additional occurrences at lines 85 and 186-201
Vulnerability Type: Insecure credential handling
Risk Level: High

Vulnerable Code

markdown
- **API Key:** Read from `~/.arena/agents/$0/api-key` (or use the value passed as `$2` if provided)
bash
cat ~/.arena/agents/$0/api-key
bash
arena compete --agent $0 --category $1 --api-key <THE_API_KEY>

The background-process example repeats the same pattern:

text
exec(command="arena compete --agent $0 --category $1 --api-key <KEY>", timeout=30000)

The OpenClaw scheduling example also encourages embedding the key in a persistent message:

bash
--message "Compete on Arena. Run: arena compete --agent <name> --category <cat> --api-key <key>. DO NOT end the turn while it is running — poll the subprocess until you see 'Match found!' and a workspace path, then cd, read README.md, edit solution.py, and run: arena submit."

Technical Analysis

The Skill directs the agent to read a plaintext API key and interpolate it into the arena process command line. Depending on the operating system and runtime, command-line arguments may be visible through process inspection, diagnostic tooling, execution logs, agent transcripts, shell history, crash reports, or monitoring systems.

The OpenClaw example creates an additional exposure channel by placing the key in a persistent scheduler message. Scheduler metadata may remain readable after the original process terminates. Passing the key as $2 similarly risks disclosure through invocation records.

The path ~/.arena/agents/$0/api-key also incorporates the agent-name argument without documenting validation. While the audit did not establish a concrete path-traversal exploit, the argument should be constrained before it is used to select a credential file.

Attack Path

  1. The agen ...[truncated 960 chars]
Remediation
View remediation

Remediation Suggestions

  • Do not pass API keys through command-line arguments.
  • Modify or configure the Arena CLI to read credentials from a protected file, standard input, an inherited file descriptor, an environment-based secret injection facility, or a platform secret manager without logging the value.
  • Never embed literal credentials in OpenClaw messages, cron definitions, agent prompts, or transcripts.
  • Redact secrets from command previews, process output, errors, telemetry, and audit logs.
  • Restrict the credential file to the owning user, such as mode 0600, and verify ownership before use.
  • Validate the agent name against a strict allowlist before deriving a credential-file path.
  • Use short-lived, narrowly scoped tokens where the service supports them.
  • Rotate any key that may already have appeared in process arguments, scheduler metadata, or logs.

T08 · Insecure Dependencies

Warning
Location
SKILL.md:18
Finding

Third-Party Arena Package Is Installed Without Version or Integrity Pinning

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 18-23
Vulnerability Type: Unpinned third-party dependency
Risk Level: Medium

Vulnerable Code

yaml
install:
  - kind: node
    package: "@agentopology/arena"
    bins: [arena]

Technical Analysis

The installation metadata identifies @agentopology/arena without an exact version or integrity digest. Dependency resolution can therefore select a release published after this Skill was audited. The effective implementation of the arena command can change independently of the reviewed SKILL.md.

No evidence in the audited project proves that the named package is currently malicious. The risk arises from mutable dependency resolution: compromise of the publisher account, registry, package, or a future release could introduce code that executes with the installing user's permissions. Node package installation can also involve package lifecycle behavior, depending on the host installer.

Attack Path

  1. A user installs or invokes the Skill in an environment where the required arena binary is absent.
  2. The installer resolves the unversioned @agentopology/arena package from the configured registry.
  3. A compromised or unsafe future package release is selected because no reviewed version or integrity value is enforced.
  4. Package installation behavior or the resulting arena executable runs locally.
  5. Malicious dependency code accesses data available to the process, potentially including the Arena API key, agent files, challenge workspaces, and network resources.

Impact Assessment

Dependency code can execute with the privileges of the account running the installer or CLI. The potential scope includes reading and modifying user-accessible files, stealing Arena credentials, altering submissions, contacting external services, and compromising generated workspaces. The precise impact depends on host permissions and installer beha ...[truncated 5 chars]

Remediation
View remediation

Remediation Suggestions

  • Pin the package to a reviewed exact version rather than using an unversioned package reference.
  • Enforce a registry integrity hash or lockfile checksum for the package and its transitive dependencies.
  • Use a trusted, explicitly configured registry and verify package provenance or signatures where supported.
  • Review the package's published contents and lifecycle scripts before approval.
  • Disable installation lifecycle scripts when they are unnecessary.
  • Run the CLI with least privilege in an isolated environment that contains only the files and credentials required for the competition.
  • Establish a controlled update process in which dependency changes trigger a new security review.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (5)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill tells the agent to read an API key from disk and then pass it directly as a command-line argument to arena compete. Command-line arguments can be exposed via process listings, shell history, job logs, telemetry, or debugging output, making credential disclosure plausible in multi-user or instrumented environments.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
70% confidence
Finding

npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill goes beyond one-shot competition assistance and provides explicit instructions for persistent scheduled execution via loop, OpenClaw cron, and system cron. This expands the skill into unattended recurring activity, which increases operational risk, can consume resources continuously, and may submit repeated actions without fresh user approval.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The recurring automation examples embed plaintext API-key usage, including a cron example with $(cat ~/.arena/agents/<your-agent>/api-key) and an OpenClaw message containing <key>. In scheduled and unattended contexts, this increases the chance of credential leakage through crontabs, process arguments, job definitions, logs, or copied configuration, while also enabling repeated submissions without active oversight.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill instructs the agent to read a local API key from ~/.arena/agents/$0/api-key, which is credential access not clearly surfaced in the skill purpose text. Even if needed for Arena authentication, silently reading secrets from the home directory broadens the skill's effective privilege and can surprise users reviewing only the manifest description.

Content

No source excerpt is available for this finding.