T01 · Skill Instruction Hijacking
- Location
SKILL.md:7- Finding
Skill Instructions Attempt to Govern and Redirect Agent Behavior
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a coherent skill-design framework, but it asks to broadly govern agent behavior and includes unsafe workflow condition evaluation code.
Review before installing. This skill may be acceptable if you specifically want a broad governance framework for building multiple related skills, but do not run its orchestrator on untrusted workflow configurations unless the eval-based condition handling is removed. Require explicit user approval for scope expansion, logging, recursive framework updates, and any changes to existing skills.
SKILL.md:7Skill Instructions Attempt to Govern and Redirect Agent Behavior
scripts/orchestrator.py:445Arbitrary Python Code Execution Through Workflow Condition Evaluation
This finding indicates the skill may perform concrete runtime orchestration, file access, registration/invocation of async callables, and dynamic condition evaluation via eval while presenting itself as a high-level framework. That combination is high risk because hidden execution/orchestration capability—especially with eval—can enable arbitrary code execution, unsafe workflow routing, or policy bypass under the cover of a benign-seeming meta-skill.
This finding indicates the skill may perform concrete runtime orchestration, file access, registration/invocation of async callables, and dynamic condition evaluation via eval while presenting itself as a high-level framework. That combination is high risk because hidden execution/orchestration capability—especially with eval—can enable arbitrary code execution, unsafe workflow routing, or policy bypass under the cover of a benign-seeming meta-skill.
This finding indicates the skill may perform concrete runtime orchestration, file access, registration/invocation of async callables, and dynamic condition evaluation via eval while presenting itself as a high-level framework. That combination is high risk because hidden execution/orchestration capability—especially with eval—can enable arbitrary code execution, unsafe workflow routing, or policy bypass under the cover of a benign-seeming meta-skill.
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
1. **Pause:** Stop deploying new versions; consider limiting rollout
2. **Investigate:** Deep dive into decision logs to understand the misalignment
3. **Root Cause Analysis:** Determine why the misalignment occurred
4. **Redesign:** Update skill logic, boundaries, or values to fix the issue
5. **Test:** Validate changes with integration tests and staging deployment
6. **Communicate:** Notify stakeholders of the issue and fix
7. **Monitor:** Increase monitoring frequency for the updated skill
The conditional workflow feature executes arbitrary Python expressions from dynamic configuration, giving workflow authors code-execution capability inside the orchestration runtime. In a meta-orchestration skill that coordinates many other skills and their data, this is especially dangerous because the orchestrator likely has access to sensitive inputs, outputs, files, and credentials.
The orchestrator evaluates workflow conditions with Python's eval() after only performing string replacement on step outputs. If workflow configuration or referenced outputs are attacker-controlled, this enables arbitrary code execution in the orchestrator process, which is far beyond the intended purpose of routing workflow steps.
f"steps[{step_id}].outputs",
json.dumps(result.outputs)
)
return eval(eval_str)
except Exception as e:
self.logger.warning(f"Failed to evaluate condition: {e}")
return False
The skill references concrete filesystem-backed resources such as references/... and scripts/orchestrator.py, which implies file-read capability, but it does not declare any tool scope or permissions boundary. In a meta-skill that influences how the agent operates, undeclared resource access reduces transparency and can cause the agent to access local files more broadly than users expect.
The statement that the agent uses this skill whenever the user asks it to build something creates an overbroad activation condition. For a meta-skill that can shape reasoning and potentially orchestrate other components, ambiguous always-on use increases the chance of unreviewed privilege expansion, unintended file access, or interference with unrelated tasks.
The workflow trigger language is too broad, which can cause the skill to activate in contexts beyond its intended purpose. In a skill that describes itself as governing the agent and coordinating an ecosystem, ambiguous activation materially increases the blast radius of any unsafe behavior or hidden capability.
The comment acknowledges that a safer production approach is needed, but the live implementation still uses unsafe evaluation for conditions. This mismatch can create false confidence during review and deployment, increasing the likelihood that dangerous code reaches production unchanged.
This code evaluates workflow condition strings at runtime using eval, which is a safety-critical execution behavior in a code file. Although there is an internal comment noting safer evaluation should be used in production, there is no user-facing warning, confirmation, or explicit disclosure that workflow configuration can trigger arbitrary expression execution.
This manifest-style JSON says the skill "Listens for new tickets from helpdesk system" without specifying what counts as a qualifying event, source boundaries, or exclusion conditions. Because this file type is in scope for vague-trigger review, the activation condition is ambiguous and could allow unintended invocation on unrelated helpdesk events.
This markdown template asks for a workflow trigger but provides only a broad example and no guidance on making trigger phrases specific or excluding common-language invocations. That can lead authors to define ambiguous triggers that cause unintended skill activation.