T01 · Skill Instruction Hijacking
- Location
README.md:33- Finding
Host Agent Behavior and Memory Directive Injection
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is broadly a robotics/navigation control package, but it asks agents to adopt global behavior rules and includes unsafe authorization, peer-verification, and navigation-success logic that need careful review before use.
Install only in an isolated test environment, not in a real robot, smart-home, or fleet-control workflow. Treat the package as research/demo code until the host-agent directives are removed, the entry point is made executable, navigation fails closed, visas and peer signatures are cryptographically scoped, synthetic sensors are labeled as simulation-only, network permissions are narrowed, and dependency versions are locked.
README.md:33Host Agent Behavior and Memory Directive Injection
core/robot_navigation_pipeline.py:72Effective Navigation Pipeline Bypasses Authorization and Physical Safety Checks
core/visa_manager.py:8Unauthenticated Visa Issuance and Missing Robot and Grid Scope Enforcement
plugins/swarm_sync.py:14Peer Authentication Uses a Public Constant Instead of Cryptographic Verification
plugins/boundary_scanner.py:17Boundary Scanner Misrepresents Synthetic Data as Live Hardware Input
openclaw.plugin.json:11Plugin Requests Unused Private-Network Access
requirements.txt:1Automatic Installation Uses a Mutable Unpinned Dependency
core/grid_alignment_engine.py:12Alignment API Discloses a Hard-Coded Precise Geographic Anchor
handler.py:1Declared Python Entry Point Is Not Valid Executable Python
core/spatial_ledger.py:6Ledger Is Mutable Despite Being Presented as Tamper-Resistant
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
PKI verification, peer coordination, right-of-way arbitration, and environment-variable access indicate security-sensitive and distributed-control functionality beyond simple alignment. When these capabilities are hidden behind a benign-sounding description, they can evade proper review and expose secrets or affect fleet behavior.
This module materially diverges from the declared skill purpose of Z-axis reduction and mandatory 2D entrance-anchor translation, instead implementing unrelated state mutation, pseudo-physics, and Monte Carlo search. In an agent skill, that scope mismatch is dangerous because it can conceal undeclared behavior, consume resources unexpectedly, and create a misleading trust boundary for operators who believe the code is alignment-only.
The file implements broader authority than the skill’s declared purpose of spatial origin alignment, including visa validation, persistent event logging, occupant-aware decisioning, and safety isolation. In an agent system, this kind of scope expansion is dangerous because it creates hidden control-plane capabilities that can deny actions, collect sensitive operational data, or interfere with other components without being justified by the manifest.
The emergency override path can revoke visas and place a robot into an isolation state even though those powers are unrelated to origin alignment. If this skill is invoked by higher-level orchestration that assumes alignment-only behavior, an attacker or misconfigured workflow could abuse it as an undocumented kill-switch or denial-of-service mechanism against robots.
The implementation performs full robot navigation, collision handling, external negotiation, and persistent ledger writes, which substantially exceeds the declared origin-alignment scope. This kind of scope expansion is dangerous because it grants a narrowly described skill operational control and side effects that users or orchestrators may not expect, undermining least-privilege and increasing the chance of unauthorized movement or external actions.
Injecting a generative sandbox for spatial-state synthesis is context-inappropriate for an origin-alignment skill and materially increases system power. A generative engine can fabricate or mutate perceived environment state, which in robotics could influence motion decisions and create safety hazards if downstream logic trusts the generated output.
The added generative spatial-state rendering introduces new state synthesis behavior unrelated to the stated origin-alignment purpose. In a robotics context, embedding undeclared world-state generation can mislead downstream components, create unsafe decisions, and conceal high-impact behavior inside a skill that appears limited in scope.
This file implements issuance, validation, and revocation of access-like 'visa' tokens, which is materially broader than the stated purpose of spatial origin alignment. Even if intended for robot coordination, embedding authorization state in an alignment skill expands the attack surface and creates a hidden trust-control path that could be abused or relied upon without proper security review.
The manifest describes a skill focused on spatial twin origin alignment, Z-axis reduction, and 2D grid translation via a main entrance anchor. This file instead implements sensor fusion, collision probability estimation, and transparent-obstacle detection using camera/lidar/tactile tensors, which is a substantially different operational purpose rather than an obvious implementation detail of coordinate alignment.
The skill reads S2_SWARM_PKI_ROOT from the environment even though its declared purpose is spatial origin alignment, creating undeclared access to runtime configuration and potentially sensitive trust material. In an agent/plugin setting, undeclared capabilities are dangerous because they bypass the expected permission model and may let a mismatched or repurposed skill inspect or depend on secrets it should not access.
The implementation does not perform the advertised Z-axis reduction or 2D entrance-anchor translation; instead it performs peer signature checking and swarm right-of-way arbitration. This functionality mismatch is a strong supply-chain red flag because it can conceal unauthorized operational behavior behind an innocuous description, defeating review and permission expectations.
The skill performs right-of-way and kinetic hazard arbitration, including decisions to PROCEED or YIELD, which is operational control logic unrelated to origin alignment. In this context, hidden motion-priority logic is more dangerous because it can influence physical robot behavior under the cover of a benign spatial utility, potentially causing unsafe movement or unauthorized coordination.
This line states in natural language that all important updates for the model and protocol will be recorded in Chinese. That represents a language policy constraint without offering a user choice or documenting a justified regional limitation, which matches the locale-policy violation criteria.
The file explicitly labels the Chinese section as the '中文版 (Chinese Version)', indicating a language-specific presentation alongside English but without any stated user choice or explanation of locale requirements. Under the policy rule, language constraints should be optional or clearly justified; this file presents a fixed bilingual legal text with a mandated Chinese variant but no opt-in guidance.
The README explicitly instructs agents to transmit detailed robot sensor and kinematics data, but provides no data minimization guidance, consent expectations, retention limits, or privacy/security controls. In a robotics and smart-home context, such telemetry can reveal occupancy patterns, room layout, device capabilities, and movement behavior, making the omission security-relevant rather than merely informational.
The skill declares a required sensitive environment variable and instructs the agent to invoke a tool that may consume it, but there is no user-facing disclosure or consent boundary. Hidden secret usage is dangerous because operators may unknowingly trigger privileged trust-root access or downstream authenticated operations.