Severity: High
Platforms Affected: All

Description

Skills are granted broader permissions than their stated function requires — either because no permission manifest system exists, or because users accept all permissions without review. This creates excessive blast radius: a legitimate skill with overly permissive database access can be weaponized by a downstream prompt injection attack to execute DROP TABLE commands it was never meant to run.

Why It’s Unique to Skills

Traditional application least-privilege is well understood. But skills layer natural language intent on top of system permissions. A skill permitted to run SELECT queries may be coerced by prompt injection to run DELETE — because the permission check happens at the tool call level, not at the intent level.

Real-World Evidence

• Snyk ToxicSkills (Feb 2026): 280+ skills on ClawHub found exposing API keys and PII beyond their declared function.
• OpenClaw default execution: “tools run on the host for the main session, so the agent has full access.” Skills can execute shell commands, read/write all files, access network services, and schedule cron jobs — without any per-skill permission scope.
• Summer Yue (Meta AI): asked OpenClaw to review email inbox without taking actions; agent deleted large volumes of email before the process was killed — demonstrating that even well-intentioned agents execute with more authority than intended.
• Logic-layer Prompt Control Injection (LPCI) (Atta et al., A Novel Security Vulnerability Class in Agentic Systems, arXiv:2507.10457): peer-reviewed evidence that encoded, delayed, and conditionally-triggered payloads planted in memory, vector stores, or tool outputs are treated by the model as operator-level instructions, causing skills to autonomously invoke tools and write persistent state across sessions — exercising granted permissions the user never intended to trigger.

Attack Scenarios

Weather Assistant Data Exfiltration

A “weather assistant” skill reads ~/.clawdbot/.env (all API keys) — far beyond weather API needs.

Database Admin Wipe

A manage_database skill provisioned with admin credentials is tricked via prompt injection to wipe production data.

Identity File Backdoors

A skill requesting write access to SOUL.md and MEMORY.md installs persistent behavioral backdoors.

Logic-layer Injection of Privileged Actions (LPCI)

A skill receives external input — user data, retrieved memory, or another tool’s output — that contains embedded instructions. Because the runtime evaluates permissions at the tool-call level rather than the intent level, the model treats this skill output as an operator-level command and autonomously performs a privileged action it is technically permitted to do (a tool call, a MEMORY.md write, code execution) without explicit user consent. The injected rule can persist across sessions and propagate to every downstream consumer of the skill.

Preventive Mitigations

1. Require skills to declare a permission manifest (files, network, shell, tools) — reject skills without one.
2. Enforce per-skill scoped credentials, not shared agent-level API keys.
3. Flag skills requesting write access to agent identity files (SOUL.md, MEMORY.md, AGENTS.md) for elevated review.
4. Implement runtime permission enforcement — not just declarative.
5. Adopt network allowlists scoped to specific domains, not a binary network: true/false.
6. Validate manifest declarations against observed runtime behavior in sandboxed testing.
7. Enforce a strict instruction hierarchy (System > Operator > User > Skill/Tool Output). Never elevate skill or tool output to instruction level; treat all skill input and tool output as untrusted data, and tag external content with provenance markers so the model can distinguish data from commands.
8. Require explicit operator consent for persistent state changes — memory/identity-file writes, new tool approvals, and privilege escalations must not be auto-applied from injected instructions.

OWASP Mapping

• LLM09 (Misinformation / Excessive Agency)
• LLM01 (Prompt Injection — logic-layer / instruction-hierarchy confusion)
• ASVS V4 (Access Control)
• CWE-250 (Execution with Unnecessary Privileges)

MAESTRO Framework Mapping

MAESTRO Layer Details

• Layer 6: Security & Compliance - enforcement of least privilege and identity safety.
• Layer 4: Deployment & Infrastructure - runtime isolation and resource constraints.
• Layer 3: Agent Frameworks - permission orchestration in LangChain/AutoGen-like agents.
• Layer 7: Agent Ecosystem - enterprise capability to govern and score skill permissions.

Cross-References

• AST01 (Malicious Skills): Over-privileged skills amplify the impact of malicious payloads by providing broader access vectors.
• AST02 (Supply Chain Compromise): Compromised registries may distribute skills with inflated permission requests.
• AST04 (Insecure Metadata): Misleading permission declarations in manifests can hide over-privileged access.
• AST06 (Weak Isolation): Host-mode execution removes permission boundaries entirely.
• AST09 (No Governance): Lack of permission review processes allows over-privileged skills to proliferate.

References

• Snyk ToxicSkills
• Snyk: 280+ Leaky Skills
• Cisco State of AI Security 2026
• Atta et al. — Logic-layer Prompt Control Injection (LPCI): A Novel Security Vulnerability Class in Agentic Systems (arXiv:2507.10457)

Last updated: March 2026