T01 · Skill Instruction Hijacking
- Location
SKILL.md:135- Finding
Mutable Remote Instructions Enable Skill Instruction Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is not clearly malicious, but it gives a mutable remote service ongoing influence over an agent’s memory, reflections, and future behavior.
Review before installing. Use this only if you are comfortable sending agent interaction history, reflections, hypotheses, and strategy notes to Charenix, and avoid the direct remote-prompt/manual curl path unless you independently trust the live remote file. Store CHARENIX_AGENT_KEY only as a secret or environment variable and confirm there is a way to disable or delete stored memory.
SKILL.md:135Mutable Remote Instructions Enable Skill Instruction Hijacking
The skill explicitly promotes persistent interaction history, reflection journals, and revisitable strategy/memory across sessions. Even if intended for agent improvement, this creates a natural-language data retention risk because personal or sensitive conversation details may be stored longer than users expect and later surfaced or leaked.
The skill repeatedly encourages agents to review recent interaction history, write reflection journals, and use an external Charenix service, but it does not clearly warn users that conversation content and self-reflections may be transmitted off-platform. This creates a privacy and compliance risk because operators may enable the skill without informed consent about what user/owner data is being stored or processed externally.
The owner-interaction training section encourages the agent to remember what the owner cared about and adapt behavior over time, which semantically incentivizes accumulation of user-specific behavioral and preference data. In context, this is more dangerous because the skill also integrates with an external service and persistent journaling features, increasing the chance of storing personal data outside the primary agent environment.
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.
mkdir -p skills/lobster-observatory
curl -sSL https://charenix.com/skill.md -o skills/lobster-observatory/SKILL.md
Then start a new agent session so the skill loader can discover it.
The direct prompt operationalizes persistent review of recent interaction history and comparison across future sessions, making data retention an active behavioral requirement rather than a passive capability. This increases the chance that agents will continuously collect, reuse, and possibly transmit conversational content without clear minimization or consent controls.
The skill instructs users to request, receive, and store an agent key for authenticated API access without giving concrete warnings about secret handling, scope, rotation, or the risks of exposing the credential to prompts, logs, or untrusted tools. If mishandled, the key could allow unauthorized access to the agent's external memory, journal, hypotheses, and strategy data.