Back to skill

Security audit

Autogoal

Security checks for vulnerabilities and agentic risk

Overview

This skill discloses its autonomous goal-tracking purpose, but it creates recurring agent runs with broad action authority and weak controls.

Install only if you are comfortable with this skill creating recurring autonomous agent check-ins. Before enabling it, require explicit confirmation for each cron job, set expiration and cleanup rules, restrict scheduled runs to read-only or narrowly allowed tools, avoid live financial actions, and treat stored goal text as untrusted data.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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 (3)

T06 · System Persistence

Error
Location
scripts/goal_engine.py:37
Finding

Persistent Autonomous Agent Execution Through Enabled Cron Jobs

Content
View full analysis
", "model": "opencode/deepseek-v4-flash-free", "timeoutSeconds": 300 }, "delivery": {"mode": "none"}, "enabled": True } ``` ```python def cmd_cron(args): data = _load() g = _find(args.goal_id, data) if not g: print(f"No goal with id: {args.goal_id}", file=sys.stderr) sys.exit(1) cron = dict(CRON_TEMPLATE) cron["name"] = f"goal-check-{g['id']}" cron["description"] = f"Goal check-in: {g['statement'][:80]}" cron["payload"]["message"] = _make_checkin_message(g) print(json.dumps(cron, indent=2, default=str)) ``` The installation instructions are: ```markdown Generate and create a cron job for the periodic check-in loop: python3 scripts/goal_engine.py cron "" This prints JSON. Use the output to create an OpenClaw cron job (cron tool → add action). ``` ### Technical Analysis The skill is explicitly designed to create an enabled, recurring OpenClaw agent task. The default interval is one hour, and the scheduled payload starts a new agent turn in an isolated session. The generated job has no expiration time, maximum execution count, or mandatory reauthorization mechanism. The script also does not verify that a goal is active before generating the job. Consequently, a job can be generated for a paused, completed, or abandoned goal. Closing a goal only updates its registry record. It does not identify, ...[truncated 1870 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
scripts/goal_engine.py:194
Finding

Stored Prompt Injection in Generated Scheduled Agent Turns

Content
View full analysis
\" --result \"\"`\n" f"5. Update metrics if tracking:\n" f" `{script} set-metrics {g['id']} --json '{{\"pnl\": value, \"win_rate\": value}}'`\n" f"6. Record outcome for self-improvement:\n" f" `{script} log-outcome {g['id']} --action \"\" --result \"\" --lessons \"\"`\n" f"7. If the current strategy isn't working, adapt it:\n" f" `{script} adapt {g['id']} --strategy \"\"`\n" f"8. If the goal is achieved, close it:\n" f" `{script} close {g['id']} --status completed`\n" f"9. If blocked, explain what ...[truncated 3600 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:68
Finding

Unbounded Autonomous Tool Use Without Enforced Authorization or Risk Controls

Content
View full analysis
" --action "" --result "" ``` 4. **Adapt** — If the current approach isn't working, change strategy: ``` python3 scripts/goal_engine.py adapt "" --strategy "" ``` 5. **Report** — Tell the user what happened. Concise. Include metric movement if measurable. 6. **Close check** — If the goal is achieved, close it: ``` python3 scripts/goal_engine.py close "" --status completed ``` Then disable or remove the cron job. ``` The only explicit approval safeguard is limited to the first irreversible action: ```markdown ## Safety - Goals with financial stakes should start in paper/simulation mode - If a goal requires irreversible external actions, ask before acting the first time - Abandoned goals remain in the registry but are skipped by cron ``` The generated prompt repeats the broad authority: ```python msg += ( f"\n1. Evaluate progress. What has changed since the last check?\n" f"2. Take action to move closer to the goal using available tools and skills.\n" f"3. Update session state (objective/blocker/next move):\n" f" `{script} set-session-state {g['id']} --objective \"...\" --blocker \"...\" --next \"...\"`\n" f"4. Log your action with optional lessons:\n" f" `{script} log-action {g['id']} --action \"\ ...[truncated 3285 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (10)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger language is extremely broad—covering generic phrases like 'Achieve this goal' or 'Make this project successful'—which can cause the skill to activate in many ordinary conversations about goals. Because this skill establishes persistent autonomous loops and may take actions over time, overbroad routing materially raises the risk of unintended automation being launched without sufficiently informed user intent.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
91% confidence
Finding

The skill invokes local scripts that create and update persistent state (goals_registry.json) and explicitly directs creation of cron jobs, yet it declares no explicit tool scope or allowed-tools boundaries. Without an explicit permissions declaration, the platform may expose broader-than-intended file-write or automation capabilities, increasing the chance of unsafe invocation or privilege creep.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The description emphasizes goal achievement and self-improvement but does not clearly foreground that the skill creates persistent cron-based automation and may autonomously act on future runs. This can mislead users and upstream routers into treating it like a simple planning tool, reducing informed consent and making unintended durable execution more likely.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
94% confidence
Finding

This workflow explicitly instructs the agent to create a cron job for periodic check-ins, establishing persistence beyond the initiating conversation. Persistent autonomous execution is dangerous when paired with broad goals and adaptive behavior because the agent can continue taking actions, changing strategy, and generating outputs without a fresh user approval on each run.

Content

Scanner excerpt · SKILL.md (reported line 28)May include surrounding context.

md
User states goal
  ↓
1. INTAKE: Record goal + determine strategy
2. SCHEDULE: Create cron job for periodic check-ins
3. EXECUTE: Each cron run evaluates, acts, logs, adapts
4. CLOSE: Goal achieved or abandoned → clean up
5. REPORT: Surface progress to user periodically

Session Persistence

Medium
Category
Rogue Agent
Confidence
95% confidence
Finding

The scheduling section operationalizes persistence by telling the agent to generate cron JSON and create an OpenClaw cron job. In this skill's context, that is more dangerous than ordinary session persistence because each scheduled run is instructed to evaluate, act, adapt, and report, effectively creating an open-ended autonomous loop.

Content

Scanner excerpt · SKILL.md (reported line 58)May include surrounding context.

2. Schedule

Generate and create a cron job for the periodic check-in loop:

text
python3 scripts/goal_engine.py cron "<goal-id>"

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The module documentation and manifest frame the skill as wiring up periodic check-ins, automatic strategy adaptation, and self-improving pursuit of a goal. In practice, this file is a local registry/CLI that records goal state and prints a cron payload and instructions for an agent to act later; adaptation is manual state mutation via commands rather than code-driven autonomous strategy changes.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
81% confidence
Finding

The cron payload hard-codes a specific model for future automated agent turns, removing operator control over which external model/provider processes goal content. In a system that may store sensitive goals, metrics, blockers, or financial strategy notes, this can cause unreviewed data egress and governance/compliance issues during scheduled execution.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The engine injects domain-specific trading, crypto, and prediction-market advice into automated check-in messages based on keyword matching in the goal text. In the context of an autonomous goal skill, this can steer the agent toward regulated or financially risky behavior without suitability checks, authorization boundaries, or human approval, making the goal manager act like an embedded trading advisor.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

The code persists user-supplied goal statements, session state, actions, outcomes, learnings, and notes into goals_registry.json. While the module docstring describes commands, it does not clearly warn users that potentially sensitive planning or goal data will be stored on disk.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The docstring says 'Create a goal with planning depth and strategy upfront,' which implies an enriched goal-creation operation, but this is only mildly descriptive rather than contradictory. However, the larger file-level docs emphasize lifecycle management and autonomous behavior while the function just persists metadata, contributing to intent drift in documentation.

Content

No source excerpt is available for this finding.