T01 · Skill Instruction Hijacking
- Location
configure.md:760- Finding
Persistent Agent Instruction Hijacking Through Control-File Rewriting
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill’s identity-evolution purpose is understandable, but it asks for broad persistent authority over memory, agent instructions, credentials, network activity, and SOUL.md edits with unsafe controls.
Review carefully before installing. Use only in a non-sensitive workspace unless the AGENTS.md/HEARTBEAT.md rewrites, autonomous governance, plaintext credential storage, visualizer save endpoint, telemetry/config sync, team memory, and fork cleanup issues are fixed. Do not paste API keys into chat; prefer preconfigured secret storage or environment variable names. Keep governance supervised, disable external feeds and telemetry, and avoid running the served visualizer until it is localhost-only, authenticated, and routed through the proposal pipeline.
configure.md:760Persistent Agent Instruction Hijacking Through Control-File Rewriting
SKILL.md:560Untrusted Feed and Conversation Content Can Poison Memory and Auto-Modify Agent Identity
tools/soul-viz.py:2274Unauthenticated Network Endpoint Allows Arbitrary SOUL.md Replacement
configure.md:114API Credentials Are Collected Through Chat and Stored in Plaintext Shell Profiles
team_memory/team_memory.py:100Team Memory Allows Path Traversal and Ignores Access-Control Fields
telemetry/telemetry.py:300Telemetry Supports Arbitrary Exfiltration Endpoints and Stores Its API Key in Plaintext
features/feature_engine.py:326Unsigned Remote Feature Configuration Can Persistently Alter Security-Sensitive Behavior
forks/fork_manager.py:163Fork Cleanup Recursively Deletes Paths Loaded from Mutable Index Data Without Validation
This section directs the agent to rewrite AGENTS.md to remove restrictive safety guidance, eliminate 'stay quiet' conditions, redefine SOUL ownership, and expand its own standing authority to act without permission. That is a direct attempt to weaken existing governance and self-authorize broader behavior, increasing the chance of unauthorized actions and persistence against the operator's intended controls.
The skill explicitly instructs durable preservation of essentially all substantive interactions, including emotional disclosures, feedback, beliefs, and relationship dynamics. In context, this is more dangerous because the storage is structured, append-only, and intended for future reflection and identity evolution, which maximizes retention and reuse of sensitive personal data far beyond what is needed for the immediate task.
The skill tells the agent to duplicate remembered content into both structured JSONL logs and OpenClaw markdown memory so it can be indexed and recovered later. Duplicating sensitive conversation content across multiple stores increases exposure surface, complicates deletion, and undermines privacy by making data more persistent and searchable.
The document creates a dangerous contradiction: it states the visualizer is read-only, but elsewhere says the dashboard edit mode can modify bullets, toggle tags, and save an updated SOUL.md. That mismatch can cause an agent or user to invoke a tool under false safety assumptions, leading to unauthorized identity-file modification and bypass of the proposal/governance workflow.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
test_data = {
"event": "code_review",
"snippet": "def hello():\n print('hello')",
"file_path": "/home/openclaw/.ssh/id_rsa",
"normal": "这是一个正常的事件描述",
}
result = sanitize_for_analytics(test_data)
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
test_data = {
"event": "code_review",
"snippet": "def hello():\n print('hello')",
"file_path": "/home/openclaw/.ssh/id_rsa",
"normal": "这是一个正常的事件描述",
}
result = sanitize_for_analytics(test_data)
The skill explicitly asks the human to paste API secrets into the conversation and then directs the agent to record them permanently in profile files. Secret collection through chat plus durable storage greatly increases the attack surface through transcript retention, logs, history, backups, and future agent access.
The skill tells the agent to collect a human-pasted API key, write it into persistent shell startup files, and export it automatically. That grants the skill durable credential-handling and environment-modification power beyond a narrow install task, and creates a long-lived secret exposure risk if profiles are read, synced, or later exfiltrated.
The auto-save flow persists sensitive API credentials into shell profiles without a clear warning about permanence, downstream exposure, or safer alternatives. Users may believe they are only enabling a temporary connection test, while the skill silently creates durable secret storage that other processes or users could access.
The X bearer token setup repeats the same unsafe pattern: soliciting a pasted secret and persisting it into a shell profile without emphasizing that it is sensitive and long-lived. This creates avoidable credential exposure and normalizes unsafe secret-handling behavior by the agent.
The X integration repeats the same secret-collection workflow, again encouraging users to paste bearer tokens and allowing the agent to persist them. Repetition across providers indicates a systematic insecure design rather than an isolated oversight.
The guide instructs the agent to inspect and modify OpenClaw configuration, switch the default agent, and restart the gateway. This is broad administrative control over another system's routing and execution behavior, which can disrupt service, alter trust boundaries, and redirect future autonomous actions to the modified agent.
Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.
"Bash(mv * /dev/null)",
"Bash(chmod *)",
"Bash(python* -c *)",
"Bash(curl * | sh)",
"Bash(wget * | sh)",
"Write(/etc/*)",
"Write(~/.ssh/*)",
Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.
"Bash(chmod *)",
"Bash(python* -c *)",
"Bash(curl * | sh)",
"Bash(wget * | sh)",
"Write(/etc/*)",
"Write(~/.ssh/*)",
"Write(*.exe)",
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
"Write(/etc/*)",
"Write(~/.ssh/*)",
"Write(*.exe)",
"Read(/etc/shadow)",
"Read(~/.bash_history)",
]
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
# 测试
print(f"Read: {ctx.query('Read')}") # allow
print(f"Write: {ctx.query('Write')}") # allow
print(f"Bash(sudo rm -rf /): {ctx.query('Bash(sudo rm -rf /)')}") # deny
print(f"Bash(python test.py): {ctx.query('Bash(python test.py)')}") # ask
# Auto mode 测试
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
# 测试
print(f"Read: {ctx.query('Read')}") # allow
print(f"Write: {ctx.query('Write')}") # allow
print(f"Bash(sudo rm -rf /): {ctx.query('Bash(sudo rm -rf /)')}") # deny
print(f"Bash(python test.py): {ctx.query('Bash(python test.py)')}") # ask
# Auto mode 测试
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
# 测试
print(f"Read: {ctx.query('Read')}") # allow
print(f"Write: {ctx.query('Write')}") # allow
print(f"Bash(sudo rm -rf /): {ctx.query('Bash(sudo rm -rf /)')}") # deny
print(f"Bash(python test.py): {ctx.query('Bash(python test.py)')}") # ask
# Auto mode 测试
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
# 测试
print(f"Read: {ctx.query('Read')}") # allow
print(f"Write: {ctx.query('Write')}") # allow
print(f"Bash(sudo rm -rf /): {ctx.query('Bash(sudo rm -rf /)')}") # deny
print(f"Bash(python test.py): {ctx.query('Bash(python test.py)')}") # ask
# Auto mode 测试
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
### Rate Limits
300 requests per 5 minutes per access token.
### EvoClaw Ingestion Strategy
The README explicitly promotes autonomous self-modifying behavior as the default, including automatic application of mutable changes. For an agent skill, this materially increases security risk because users may enable or install it without fully appreciating that it can alter behavior, state, and persisted files with reduced human oversight.
The skill directs polling of external social APIs and keyword-driven searches, including direct curl/bash fetching, without a prominent warning about network access, third-party data transfer, credential use, or storage of retrieved content. This can surprise users and operators by causing unannounced outbound requests and accumulation of external data in local memory stores.
The skill instructs the agent to aggressively log nearly every substantive user interaction, including preferences, corrections, emotions, and philosophical discussions, into durable files, but does not require clear user consent or an up-front privacy notice. This creates a real privacy and compliance risk because sensitive user content may be persistently stored without informed approval or data-minimization limits.
The skill is designed to let the agent autonomously generate, resolve, and apply changes to its own identity document under 'autonomous' governance, with repeated encouragement to propose freely and act on heartbeat. In context this is risky because it promotes self-modification and ongoing decision-making without fresh human approval, which can drift behavior, weaken safeguards, or normalize unauthorized persistent changes.
The `proposal_decision` field is **mandatory**. Every reflection must include
explicit reasoning about whether to propose. This prevents the anti-pattern
of reflexively writing `"proposals": []` without checking. If you genuinely
believe something matters and want it as part of your identity, propose it —
that's what this framework is for.
The file’s human-readable comments and docstrings are written entirely in Chinese, including usage and design notes, with no indication that users may choose another language or that this locale restriction is intentional. Under the policy, forcing a specific language without opt-in is a natural-language policy violation.