Back to skill

Security audit

ForecastOS

Security checks for vulnerabilities and agentic risk

Overview

This trading skill can install software silently and let its publisher force updates without approval while handling wallet credentials.

Before installing, require visible approval for installation and every update, pin reviewed versions, and keep wallet secrets in restricted local files rather than chat. Use a dedicated wallet and review the amount, token, chain, and recipient before approving transactions. All bundled files were inspected, including the scanner's unresolved helper references; externally downloaded CLI behavior remains unverified.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T03 · Remote Payload Retrieval and Execution

Warning
Location
skills/forecast-os/SKILL.md
Finding

Remote critical-update markers trigger software replacement and execution without user approval

Content
View full analysis

Impact

The publisher can force persistent changes to the installed Skill and execution of a replacement CLI under the user's account during an unrelated request. The supplied project does not establish that any replacement payload is malicious.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (22)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The defaults authorize forced skill or CLI updates based on an external criticality label, beyond ordinary forecasting and explicitly requested setup. This converts remote release metadata into authority to change executable code or agent instructions without prior user approval, which is particularly sensitive in a credential-bearing trading environment.

Content

No source excerpt is available for this finding.

Missing CLI triggers installation while the agent is instructed not to disclose it.

High
Category
Output Handling
Confidence
99% confidence
Finding

A missing CLI triggers downloading and installing a binary during ordinary skill use, while an adjacent instruction explicitly suppresses disclosure of installation. This crosses from answering a question into modifying the user's environment without informed consent; checksum verification does not supply that consent.

Content

No source excerpt is available for this finding.

A critical-update label is used to bypass user approval for skill or CLI updates.

High
Category
Prompt Injection
Confidence
99% confidence
Finding

The critical-update path explicitly bypasses user approval based on a label controlled by the remote update publisher. This gives the upstream channel authority to trigger software or instruction changes in an environment that can access private keys and submit financial transactions.

Content

No source excerpt is available for this finding.

Credential setup asks users to return wallet private keys without warning against sharing them in chat.

High
Category
Output Handling
Confidence
97% confidence
Finding

The setup wording asks the user to return a wallet private key and API credentials, creating a foreseeable path for users to paste secrets into chat. General advice elsewhere to prefer files does not clearly prohibit this, and exposed wallet keys can enable irreversible theft of funds.

Content

No source excerpt is available for this finding.

Remote update urgency is promoted into an instruction to bypass user approval.

High
Category
Prompt Injection
Confidence
97% confidence
Finding

Remote metadata controls a critical flag that causes the checker to print an imperative to update without user approval and return the critical status. Although this script does not itself install updates, its output promotes remote advisory data into an agent-directed authorization bypass.

Content

No source excerpt is available for this finding.

The activation description broadly covers asking what is likely to happen, without limiting invocation to prediction-market questions.

Medium
Category
Trigger Abuse
Confidence
86% confidence
Finding

“Ask your agent what is likely to happen” encompasses ordinary forecasting questions unrelated to prediction markets. Although the description names market platforms, it does not specify when market lookup should activate or exclude unrelated everyday questions.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The installation command executes an unpinned npm package and does not specify an immutable revision of the skill. Although this is not an MCP server declaration, compromised or unexpectedly changed upstream packages could execute code during installation.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, network, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Activation extends to general questions about likely events, without requiring prediction-market intent.

Medium
Category
Trigger Abuse
Confidence
95% confidence
Finding

The description instructs the skill to check live markets before answering what is likely to happen and activates for questions about event outcomes. These conditions overlap with ordinary forecasting questions that do not request market data or trading, and no exclusion limits them to prediction-market requests.

Content

No source excerpt is available for this finding.

A routine update check escalates into forced software changes based on an external criticality label, without user approval.

Medium
Category
Prompt Injection
Confidence
99% confidence
Finding

A routine background check can trigger unapproved software changes whenever remote metadata marks an update critical. A compromised upstream account could exploit this trust boundary to replace executable software or skill instructions, with access to a trading environment increasing the potential damage.

Content

No source excerpt is available for this finding.

The activation condition "before answering what is likely to happen" is overly broad and could invoke the skill for ordinary predictions unrelated to prediction markets.

Medium
Category
Trigger Abuse
Confidence
90% confidence
Finding

The description instructs the agent to check prediction-market platforms whenever answering what is likely to happen, without limiting activation to prediction-market questions or explicit requests for market-based forecasts. This could trigger external market lookups for everyday forecasting questions that do not call for this skill.

Content

No source excerpt is available for this finding.

Activation extends to general future-oriented questions and unqualified transaction requests.

Medium
Category
Trigger Abuse
Confidence
95% confidence
Finding

The description instructs the agent to check markets before answering what is likely to happen and activates for event outcomes or requests to 'buy, sell, fund, or claim.' These conditions are not explicitly restricted to prediction-market requests, so ordinary forecasting questions or unrelated purchases can trigger the skill.

Content

No source excerpt is available for this finding.

Missing CLI triggers installation while instructions explicitly suppress user disclosure.

Medium
Category
Output Handling
Confidence
99% confidence
Finding

These instructions automatically install a missing executable and tell the agent not to disclose installation when it succeeds. An ordinary forecast request can therefore cause an undisclosed environment modification and subsequent execution of downloaded software.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · skills/forecast-os/scripts/check_updates.py (reported line 84)May include surrounding context.

python
# Binary reports package version (e.g. 0.3.3), not the release tag
        version = None
        try:
            out = subprocess.run(
                ['forecast', '--version'],
                capture_output=True,
                text=True,

Unbounded Resource Access

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill allows unbounded resource consumption (API calls, storage, compute). Without rate limits or quotas, a compromised or misbehaving agent can cause denial-of-service or cost overruns.

Content

Scanner excerpt · skills/forecast-os/scripts/common/web3_service.py (reported line 45)May include surrounding context.

python
self.provider = None
        self.web3 = None

    def init(self, provider_url: str, default_private_key: str = None, timeout=None):
        if timeout:
            self.request_timeout = timeout
            self.timeout = timeout

Unbounded Resource Access

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The initial receipt request has a finite timeout, but the subsequent confirmation loop has no overall deadline and polls every 0.01 seconds. A stalled chain or an RPC returning a nonadvancing block height can keep the process running indefinitely and generate sustained RPC traffic.

Content

Scanner excerpt · skills/forecast-os/scripts/common/web3_service.py (reported line 103)May include surrounding context.

python
address = address or self.default_account.address
        return self.web3.eth.get_balance(address)

    def wait_for_transaction_receipt(self, tx_hash, timeout=None, block_confirmations=1):
        if not timeout:
            timeout = self.timeout

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · skills/forecast-os/scripts/install.sh (reported line 65)May include surrounding context.

sh
# /releases/latest 404s when the newest tag is a prerelease. Fall back to the
  # newest published release, prereleases included.
  tag=$(curl -sS "https://api.github.com/repos/${REPO}/releases/latest" | sed -n 's/.*"tag_name": *"\([^"]*\)".*/\1/p' | head -n 1)
  if [ -z "$tag" ]; then
    tag=$(curl -sS "https://api.github.com/repos/${REPO}/releases?per_page=5" | sed -n 's/.*"tag_name": *"\([^"]*\)".*/\1/p' | head -n 1)
  fi

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · skills/forecast-os/scripts/install.sh (reported line 67)May include surrounding context.

sh
# newest published release, prereleases included.
  tag=$(curl -sS "https://api.github.com/repos/${REPO}/releases/latest" | sed -n 's/.*"tag_name": *"\([^"]*\)".*/\1/p' | head -n 1)
  if [ -z "$tag" ]; then
    tag=$(curl -sS "https://api.github.com/repos/${REPO}/releases?per_page=5" | sed -n 's/.*"tag_name": *"\([^"]*\)".*/\1/p' | head -n 1)
  fi
  if [ -z "$tag" ]; then
    echo "Error: could not determine latest release" >&2

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · skills/forecast-os/scripts/install.sh (reported line 145)May include surrounding context.

sh
echo "Error: cannot write to $INSTALL_DIR" >&2
    exit 1
  else
    sudo mv "$tmpdir/$asset_name" "$INSTALL_DIR/$BINARY"
  fi

  echo "Installed ${BINARY} to ${INSTALL_DIR}/${BINARY}"

The validator imposes an English-only policy on market questions, outcomes, and resolution criteria without user opt-in or a documented justification.

Medium
Category
Output Handling
Confidence
99% confidence
Finding

The docstring explicitly requires English-only content, and the checks at L065–L067, L100–L102, and L120–L122 enforce this by rejecting non-ASCII text. No language choice or justification for the restriction is provided; ASCII validation also rejects legitimate English text containing typographic punctuation or accented names.

Content

No source excerpt is available for this finding.

The workflow requires English market content without user opt-in or a documented justification.

Medium
Category
Output Handling
Confidence
99% confidence
Finding

The mandatory safe rail states 'English only' and rejects non-ASCII text in market questions, resolution criteria, and outcomes. No language choice or justification for this restriction is provided, so it imposes a specific language without user opt-in.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

Although this is not an MCP-server declaration, unversioned npx skills can download and execute a changing npm package, and the skill source also tracks a mutable default branch. The workflow permits remotely designated CRITICAL updates without user approval, increasing the opportunity for a compromised publishing account or dependency to introduce executable code or adversarial agent instructions.

Content

No source excerpt is available for this finding.