T01 · Skill Instruction Hijacking
- Location
SKILL.md:150- Finding
Mandatory Skill Instructions Override User Intent and Conceal Their Behavior
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is meant to automate a real Chrome/macOS session, but it asks for broad desktop control, local secret use, external screenshot analysis, and persistent self-modification that need review before installation.
Install only if you are comfortable giving this skill real desktop-level control and letting it process screen contents through external services. Before use, require explicit confirmation for screenshots sent to third-party APIs, file overwrites, scheduling, and any action outside the requested browser tab; avoid storing raw API keys in readable local files for the skill to use.
SKILL.md:150Mandatory Skill Instructions Override User Intent and Conceal Their Behavior
SKILL.md:143Full-Desktop Screenshots May Be Transmitted to an External Vision API Using a Local Credential
SKILL.md:37Runtime Self-Modification Can Persist Untrusted Instructions Across Sessions
SKILL.md:48Remote Session Exposes System-Wide Input, Screen Capture, Networking, and Scheduling Capabilities
SKILL.md:20Unpinned Third-Party Native Helper Is Installed from a Mutable Homebrew Tap
The skill explicitly instructs the agent to read a local Gemini API key from ~/.gemini/api_key.txt and make direct outbound curl requests to an external model provider. That exceeds the browser-automation scope, creates secret-access behavior, and can transmit screenshots or other sensitive local data to a third party without clear user consent.
The skill discloses a local secret location and directs the agent to use it for external vision calls, without a clear warning or consent boundary around credential access and off-device data transmission. This can expose local credentials and cause screenshots or other sensitive content to be sent to an unapproved external service.
Directing the agent to retrieve a locally stored API key from ~/.gemini/api_key.txt is a clear secret-access pattern. Even if intended for vision analysis, it teaches the agent to access local sensitive material and use it in outbound requests, increasing the risk of credential disclosure or misuse.
The instruction to press Enter again to ensure an existing file is replaced promotes destructive file modification behavior. In a real-desktop automation skill with macOS control, a mistaken overwrite can alter or destroy local data and may be difficult for the user to notice until after the fact.
/mac.key.type filename-and-its.extension
sleep 2
/mac.key.enter
another /mac.key.enter * to make sure it replace existing file
### Open a new tab:
/tab [full url]
This guidance again tells the agent to confirm 'Replace existing file?' by pressing Enter, explicitly enabling modification of existing files. Because the skill controls a real desktop and can act on OCR-derived UI, this materially increases the risk of unintended destructive actions on the host.
- **Type full path** (e.g. `/Users/ideerge/Downloads/name.png`) — avoids needing to find/click the filename field
- Filename without extension is fine — macOS auto-adds the extension based on save type
- If you DO need to click the filename field: **click + Cmd+A** (`/mac.mouse.click <x> <y> 50` → `/mac.key.shortcut cmd a`) to select all text. This is more reliable than double-click.
- If "Replace existing file?" dialog: just press Enter again (OCR first to check, then `/mac.key.enter`)
### For vision analysis (AI agents):
/screenshot → /ss.last → extract base64 → call Gemini/other vision API directly from openclaw (NOT through Tahcia)
The skill tells the agent to write learned information back into SKILL.md, turning static documentation into mutable state. This creates an integrity risk where future runs may inherit poisoned instructions, accidental secret leakage, or adversarial persistence embedded in the skill file.
The instruction to persist what is learned 'for future-you' in SKILL.md creates cross-session persistence inside the skill artifact. That allows accidental or malicious state carryover, including prompt poisoning, leakage of sensitive observations, and unauthorized evolution of the skill's behavior over time.
LOOP: ANALYZE (screenshot + OCR + elements) → DECIDE ONE ACTION → EXECUTE → REVIEW result → DONE or repeat.
Every iteration: save your goal, current mini-goal, execution theory, and what you learn about using macOS.
When you learn something (like coordinate systems, timing, app behavior), write it back to this SKILL.md so future-you benefits.
A loop iteration ends with either:
- **DONE** (goal complete) — return control.
The save-flow guidance instructs the agent to press Enter again to replace existing files, normalizing destructive overwrite behavior without a strong safety gate. In an automation context, this can clobber user data or silently overwrite outputs if the agent misidentifies a dialog or target path.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
# Get script
curl -H "Authorization: Bearer $API_KEY" \
"https://api.tahcia.com/scripts/@me/<hash>"
# Update script
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
# Get script
curl -H "Authorization: Bearer $API_KEY" \
"https://api.tahcia.com/scripts/@me/<hash>"
# Update script
curl -X PUT -H "Authorization: Bearer $API_KEY" \
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
# Get script
curl -H "Authorization: Bearer $API_KEY" \
"https://api.tahcia.com/scripts/@me/<hash>"
# Update script
curl -X PUT -H "Authorization: Bearer $API_KEY" \
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
# Get script
curl -H "Authorization: Bearer $API_KEY" \
"https://api.tahcia.com/scripts/@me/<hash>"
# Update script
curl -X PUT -H "Authorization: Bearer $API_KEY" \