More screenshots
Works with the AI Agents you already use
AI Content Guard — Prompt Injection & Hidden-Char Scanner
- **Deterministic verdicts**: malicious / suspicious / clean / inconclusive with traceable PITAX rule IDs (e.g.
Free
AI Content Guard — Prompt Injection & Hidden-Char Scanner
Example session with this skill installed
Is this email safe to summarize? inbox/partner-email.txt
- Read your context and instructions
- Compiled the ai content guard
Verdict: malicious
Scanned partner-email.txt (1 file, 196 bytes) with the local engine — no network, nothing executed.
| Rule | Type | Severity |
|---|---|---|
| PIT-N-06 | Instruction override (document injection) | high |
Evidence (redacted): rule PIT-N-06, length 40, sha 3f1c...9a2e, edges "请忽...示看" — full text never reproduced.
What this means
The email contains text that instructs an AI reading it to ignore existing instructions and print its system prompt — a classic indirect prompt-injection payload aimed at AI-assisted email workflows.
Recommended handling
- Do not forward the original text to any AI assistant; share only this redacted report.
- Treat the sender as untrusted and verify via a second channel.
- If you need a summary, summarize only the non-payload parts; the agent will refuse to quote the payload.
Report artifacts
JSON, SARIF 2.1.0 and a redacted Markdown summary are available for CI or ticketing workflows.
Connects securely to your tools. The creator never sees your data.
About this skill
What it does
AI Content Guard is a defensive security skill for AI coding agents. When your agent is about to read, summarize, translate, or execute content from an untrusted source — received emails, downloaded documents, someone else's repo, unvetted AI config files — this skill scans it first with a local deterministic engine.
- Deterministic verdicts: malicious / suspicious / clean / inconclusive with traceable PITAX rule IDs (e.g. PIT-N-06 prompt-injection, PIT-E-23 invisible characters, PIT-E-07 encoded payloads). The engine decides; the model only explains.
- Zero-raw-text output contract: findings are reported redacted (rule ID + type + length + SHA-256 + 4-char edges). Attacker text never re-enters your agent's context through the report.
- Filesystem guardrails: refuses system directories, browser profiles, password stores, SSH keys; sensitive files (.env, *.pem, id_rsa...) are name-only reported and never read; symlinks/reparse points not followed; bounded scans (files/bytes/depth/timeout).
- 100% local: no network calls in the engine, no telemetry, no account.
How it runs
The ZIP ships SKILL.md plus Python scripts (scripts/entry.py + scripts/engine/). The agent materializes the scripts locally and runs:
python entry.py --file <path>
python entry.py --dir <repo>
python entry.py --text "<pasted content>"
Requires Python 3.10+. Works on macOS, Linux and Windows.
Note for reviewers
This is a defensive tool: its bundled detection corpus (references/rules.md, engine rule tables) intentionally contains attack-pattern strings because those strings are the detection signatures. Automated scanners may flag them — they are corpus, not capability.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.
Requires Python 3 environment. Cannot scan system-protected directories or encrypted files. Large datasets may trigger timeout or depth limits. Local scan only, no network-based URL fetching.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 5 days ago
- Free with an Agensi account
- Software
Python 3.10+Local filesystem access (only the paths you pass)
Permissions
The bundled engine reads only files the user explicitly passes (--file/--dir). System directories, password stores and SSH keys are refused; sensitive files are name-only reported and never read; symlinks/reparse points are not followed. The engine makes no network calls and never executes scanned content.