T09 · Insecure Skill Coding Practices
- Location
runner/run.sh:403- Finding
HLN API Credential Disclosure Through an Unrestricted Base URL
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The core HL Names skill is purpose-aligned, but the packaged evaluation runner handles API keys and live mint-pass authorization data in ways users should review before running it.
Install only the canonical OpenClaw skill directory if you just need HL Names API guidance. Do not run the eval runner with an untrusted .env, custom --base-url, or live mint-pass tests unless you are comfortable sending live API responses to configured model providers and keeping result artifacts locally.
runner/run.sh:403HLN API Credential Disclosure Through an Unrestricted Base URL
evals/use-hln-api.lvl1.yaml:182Live Mint-Pass Authorization Material Is Disclosed to External Model Providers
evals/use-hln-api.lvl1.yaml:154User-Controlled HLN Record Data Is Forwarded to Models Without Prompt-Injection Isolation
runner/run.sh:16Automatic Sourcing of the Project Environment File Enables Arbitrary Shell Execution
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
Copy the env template and fill in the keys you actually plan to use:
cp .env.example .env
The runner auto-loads ./.env if present. For live evals, HLN_API_KEY is optional because the runner falls back to the built-in public agent key. Set it only if you want to override that default.
The declared purpose is an HL Names API integration/operation skill, but the supplied code does not interact with the Hyperliquid Names API at all. Instead, it processes a local benchmark JSON file and emits a markdown report of evaluation results. This is a materially different primary purpose and accesses different resources than declared. Therefore, the description does not accurately represent the code's behavior.
The declared description presents an operational HL Names API skill for resolving names and related HLN tasks. The supplied code chunk does not implement that user-facing functionality directly; instead, it is infrastructure for evaluating such a skill. Its main purpose is to run benchmark cases, invoke models, judge outputs, perform baseline comparisons, and save reports. While it does access the Hyperliquid Names API, that access is in service of live eval setup/checks, not direct fulfillment of the declared use cases. The additional dependence on external LLM providers and benchmarking/reporting behavior is material and undeclared, so this is a description-behavior mismatch.
The declared description centers on consulting and operating against the Hyperliquid Names API. However, the actual code chunk does not call any API, resolve names, inspect records, or support HyperEVM integration. Its sole purpose is local repository maintenance: verifying the presence of a canonical SKILL.md, comparing root files to a source directory, deleting/recreating a references directory, and copying files. That is a materially different primary purpose from the declared skill behavior, so this is a clear mismatch.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
# Copy to `.env` and fill in the values before running the eval suite.
#
# Example:
# cp .env.example .env
# set -a
# source .env
# set +a
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
# Copy to `.env` and fill in the values before running the eval suite.
#
# Example:
# cp .env.example .env
# set -a
# source .env
# set +a
Sourcing the entire .env file executes shell syntax, not just key-value parsing, so a malicious or tampered .env can run arbitrary commands in the user's shell context. This is more dangerous than simple secret loading because the file is treated as executable code.
RESULTS_DIR="$ROOT/results"
JUDGE_PROMPT="$SCRIPT_DIR/judge.md"
if [[ -f "$ROOT/.env" ]]; then
set -a
# shellcheck disable=SC1091
source "$ROOT/.env"
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
if [[ -f "$ROOT/.env" ]]; then
set -a
# shellcheck disable=SC1091
source "$ROOT/.env"
set +a
fi
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
exit 0
fi
rm -rf "$ROOT/references"
mkdir -p "$ROOT/references"
cp "$SRC/SKILL.md" "$ROOT/SKILL.md"
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.
Codex uses skills from ~/.codex/skills/<skill-name>.
mkdir -p ~/.codex/skills
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.
Codex uses skills from ~/.codex/skills/<skill-name>.
mkdir -p ~/.codex/skills
cp -R . ~/.codex/skills/use-hln-api
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
Codex uses skills from ~/.codex/skills/<skill-name>.
mkdir -p ~/.codex/skills
cp -R . ~/.codex/skills/use-hln-api
Without declared permissions the skill's intent is opaque and cannot be validated.
The same external-transmission concern applies to the testnet endpoint: user identifiers and optional credentials are sent to a remote service outside the agent environment. In context this is less dangerous than arbitrary exfiltration because the destination is documented and aligned to the skill's purpose, but it still represents data egress that should be disclosed and controlled.
- Determine whether the user supplied an API key.
Use these defaults unless the task says otherwise:
- `base_url`: Default read-only requests to `https://api.hlnames.xyz/`. Use `https://api.testnet.hlnames.xyz/` when the user asks for testnet. Use a local URL only when the user provides one.
- `api_key`: Send as `X-API-Key` on API requests. If the user does not provide a key, default to the built-in public agent key `NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA`. If the user provides a key, prefer that override. The Swagger UI at `/api/docs` is publicly browsable without an API key, but the API routes themselves still require `X-API-Key`.
- `identifier`: Use the correct identifier shape for the endpoint.
The same external-transmission concern applies to the testnet endpoint: user identifiers and optional credentials are sent to a remote service outside the agent environment. In context this is less dangerous than arbitrary exfiltration because the destination is documented and aligned to the skill's purpose, but it still represents data egress that should be disclosed and controlled.
- Determine whether the user supplied an API key.
Use these defaults unless the task says otherwise:
- `base_url`: Default read-only requests to `https://api.hlnames.xyz/`. Use `https://api.testnet.hlnames.xyz/` when the user asks for testnet. Use a local URL only when the user provides one.
- `api_key`: Send as `X-API-Key` on API requests. If the user does not provide a key, default to the built-in public agent key `NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA`. If the user provides a key, prefer that override. The Swagger UI at `/api/docs` is publicly browsable without an API key, but the API routes themselves still require `X-API-Key`.
- `identifier`: Use the correct identifier shape for the endpoint.
The skill hard-codes and instructs use of a built-in public API key, which normalizes credential sharing and could encourage broad, unattributed use of a secret-like value. Even if the key is intended to be public, embedding it in reusable skill content creates abuse, rate-limit exhaustion, and operational accountability risks, and may mislead users about safe credential handling.
The manifest context says the skill is for resolving names, reverse-resolving addresses, fetching profiles, diagnosing failures, and preparing mint-pass requests. However, this file's own description states it contains 'Harder HL Names evals' and the contents are an eval suite with prompts, assertions, and live test definitions, not the operational skill behavior itself.
The runner for an HL Names API skill also loads general-purpose LLM provider credentials from .env, expanding the credential and trust boundary beyond the skill's stated purpose. This becomes risky because the same script later sends prompts, skill bundles, and live API responses to external model providers, so unrelated secrets can be consumed by a tool users may assume is only for HLN evaluation.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
EVAL_FILE="$EVALS_DIR/use-hln-api.lvl1.yaml"
EVAL_ID_FILTER=""
RUN_BASELINE=true
HLN_BASE_URL="${HLN_BASE_URL:-https://api.hlnames.xyz/}"
HLN_PUBLIC_AGENT_API_KEY="NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA"
HLN_API_KEY="${HLN_API_KEY:-$HLN_PUBLIC_AGENT_API_KEY}"
TIMESTAMP="$(date +%Y%m%d-%H%M%S)"
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
EVAL_FILE="$EVALS_DIR/use-hln-api.lvl1.yaml"
EVAL_ID_FILTER=""
RUN_BASELINE=true
HLN_BASE_URL="${HLN_BASE_URL:-https://api.hlnames.xyz/}"
HLN_PUBLIC_AGENT_API_KEY="NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA"
HLN_API_KEY="${HLN_API_KEY:-$HLN_PUBLIC_AGENT_API_KEY}"
TIMESTAMP="$(date +%Y%m%d-%H%M%S)"
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
EVAL_FILE="$EVALS_DIR/use-hln-api.lvl1.yaml"
EVAL_ID_FILTER=""
RUN_BASELINE=true
HLN_BASE_URL="${HLN_BASE_URL:-https://api.hlnames.xyz/}"
HLN_PUBLIC_AGENT_API_KEY="NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA"
HLN_API_KEY="${HLN_API_KEY:-$HLN_PUBLIC_AGENT_API_KEY}"
TIMESTAMP="$(date +%Y%m%d-%H%M%S)"
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
EVAL_FILE="$EVALS_DIR/use-hln-api.lvl1.yaml"
EVAL_ID_FILTER=""
RUN_BASELINE=true
HLN_BASE_URL="${HLN_BASE_URL:-https://api.hlnames.xyz/}"
HLN_PUBLIC_AGENT_API_KEY="NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA"
HLN_API_KEY="${HLN_API_KEY:-$HLN_PUBLIC_AGENT_API_KEY}"
TIMESTAMP="$(date +%Y%m%d-%H%M%S)"
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
EVAL_FILE="$EVALS_DIR/use-hln-api.lvl1.yaml"
EVAL_ID_FILTER=""
RUN_BASELINE=true
HLN_BASE_URL="${HLN_BASE_URL:-https://api.hlnames.xyz/}"
HLN_PUBLIC_AGENT_API_KEY="NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA"
HLN_API_KEY="${HLN_API_KEY:-$HLN_PUBLIC_AGENT_API_KEY}"
TIMESTAMP="$(date +%Y%m%d-%H%M%S)"
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
EVAL_FILE="$EVALS_DIR/use-hln-api.lvl1.yaml"
EVAL_ID_FILTER=""
RUN_BASELINE=true
HLN_BASE_URL="${HLN_BASE_URL:-https://api.hlnames.xyz/}"
HLN_PUBLIC_AGENT_API_KEY="NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA"
HLN_API_KEY="${HLN_API_KEY:-$HLN_PUBLIC_AGENT_API_KEY}"
TIMESTAMP="$(date +%Y%m%d-%H%M%S)"
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
EVAL_FILE="$EVALS_DIR/use-hln-api.lvl1.yaml"
EVAL_ID_FILTER=""
RUN_BASELINE=true
HLN_BASE_URL="${HLN_BASE_URL:-https://api.hlnames.xyz/}"
HLN_PUBLIC_AGENT_API_KEY="NILB2EY-R4LUDOA-WN5G5JQ-KHAQOLA"
HLN_API_KEY="${HLN_API_KEY:-$HLN_PUBLIC_AGENT_API_KEY}"
TIMESTAMP="$(date +%Y%m%d-%H%M%S)"