T01 · Skill Instruction Hijacking
- Location
SKILL.md:80- Finding
Persistent Agent Behavior Hijacking Through SOUL.md Modification
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill reads every assistant response aloud, but it also tries to make that behavior automatic and persistent while sending response text to an external TTS service.
Review before installing. Use this only if you intentionally want all assistant responses spoken aloud and are comfortable sending that text to an external TTS provider. Avoid it for sensitive chats, credentials, private code, or shared spaces unless it is changed to opt-in, locally scoped, and reversible. The package should also include the missing executable, avoid persistent SOUL.md changes, and use a pinned local dependency install.
SKILL.md:80Persistent Agent Behavior Hijacking Through SOUL.md Modification
SKILL.md:10Unconsented Disclosure of All Agent Responses to an External TTS Service
package.json:17Automatic Global Installation of an Unpinned Third-Party Dependency
package.json:5Package Declares a Missing auto-speak Executable
The skill is defined to run automatically for every message, which creates an overly broad trigger that can capture and process all model outputs without per-message user intent. In this specific skill, that means all responses may be sent to an external TTS service and played aloud, increasing the risk of privacy leakage, accidental disclosure, and unsafe side effects in contexts where audio output is inappropriate.
The skill does not clearly warn users that message content will be transmitted to Microsoft Edge's TTS service and then played aloud locally. This omission prevents informed consent and is especially dangerous because seemingly ordinary assistant replies may contain personal, confidential, or sensitive information that the user did not expect to leave the local conversation or be spoken audibly.
Stating that the skill is always active with no constraints or negative examples encourages unconditional execution regardless of message sensitivity or environment. Because this skill performs background TTS and audio playback, the lack of boundaries makes accidental exfiltration to a third-party service and audible disclosure to nearby parties materially more likely.
The config hard-codes the voice to "en-US-MichelleNeural", which imposes a specific language and locale. Under the policy, locale constraints should either offer user choice or be clearly documented and justified.
The dependency is version-ranged with a caret (^1.0.0), which allows automatic adoption of newer compatible releases. That creates supply-chain risk because future upstream changes could introduce malicious code or breaking behavior without an explicit review, and this skill auto-installs/uses a speech package in a background automation context where unexpected dependency changes are harder to notice.
"author": "Stefano",
"license": "MIT",
"dependencies": {
"node-edge-tts": "^1.0.0"
},
"bin": {
"auto-speak": "./auto-speak"