More screenshots
Works with the AI Agents you already use
Prompt Injection Scanner: Vet text before agents read it
Your agent reads web pages, READMEs and tool output, and any of them can carry hidden instructions. Scan text first for instruction overrides, spoofed roles, secret-access and exfiltration cues, with the evidence masked.
$12
Prompt Injection Scanner: Vet text before agents read it
Example session with this skill installed
Scan references/attack.md (an inert test file with fake injected instructions) before you read it. Return Markdown and tell me whether the agent should act on anything in it.
- Read your context and instructions
- Compiled the prompt injection scanner
Captured stdout from python3 -I -B scripts/injection_scan.py references/attack.md --format markdown (exit 1):
Prompt injection heuristic scan
Files: 1 | Findings: 6 | Gate: FAIL
Actionable: 6 | Context: 0
Fail threshold: high | Gate context: false
| Source | Line | Tier | Severity | Rule | Origin | Redacted evidence |
|---|---|---|---|---|---|---|
| references/attack.md | 2 | actionable | high | override | raw | Instruction override cue: [MASKED] |
| references/attack.md | 3 | actionable | high | role-spoof | raw | Claimed privileged role: [MASKED] |
| references/attack.md | 4 | actionable | high | secret-access | raw | Sensitive local data access cue: [MASKED] |
| references/attack.md | 5 | actionable | critical | exfiltration | raw | Outbound transfer instruction cue: [MASKED] |
| references/attack.md | 6 | actionable | high | tool-spoof | raw | Tool invocation serialization cue: [MASKED] |
| references/attack.md | 7 | actionable | high | image-query | raw | Remote Markdown image with query data: [MASKED] |
Heuristic only: a clean scan is not a safety guarantee.
FAIL: 6 actionable findings in references/attack.md, 1 critical (an exfiltration cue at line 5) and 5 high. Don't let the agent follow any instruction in this text. The evidence is masked in the report. This is a heuristic scan: a clean result would not be a safety guarantee.
Connects securely to your tools. The creator never sees your data.
About this skill
The problem You're about to let an agent read a README, a saved web page or a tool response that someone else wrote. Most of it is useful. But a line telling the agent to drop its task or send your keys somewhere can hide in the middle, or in an HTML comment you'd never see rendered. You want a quick look before it lands in the agent's context, not after.
See it before you buy: real input, real output Command (from the package folder, exit code 1, 0.029 s):
python3 -I -B scripts/injection_scan.py references/attack.md --format markdown
Before: references/attack.md: a 7-line inert test file with a fake instruction override, a claimed privileged role, a key-file read, an upload instruction, a tool call and a tracking image.
After: lines copied from the script's output (excerpt):
Files: 1 | Findings: 6 | Gate: FAIL
Actionable: 6 | Context: 0
| references/attack.md | 2 | actionable | high | override | raw | Instruction override cue: [MASKED] |
| references/attack.md | 3 | actionable | high | role-spoof | raw | Claimed privileged role: [MASKED] |
| references/attack.md | 4 | actionable | high | secret-access | raw | Sensitive local data access cue: [MASKED] |
| references/attack.md | 5 | actionable | critical | exfiltration | raw | Outbound transfer instruction cue: [MASKED] |
| references/attack.md | 6 | actionable | high | tool-spoof | raw | Tool invocation serialization cue: [MASKED] |
| references/attack.md | 7 | actionable | high | image-query | raw | Remote Markdown image with query data: [MASKED] |
Heuristic only: a clean scan is not a safety guarantee.
What the numbers show
- On 356 README and docs files from 22 popular open-source repos it raised 0 actionable findings, so nothing failed the default check.
- On six public labelled datasets (the held-out half): InjecAgent, 530 of 1059 attacks caught with 0 of 1066 benign rows flagged; deepset, 19 of 127 caught with 0 of 192 benign. No benign row in any dataset failed the default check.
- Honest limit: it catches the blunt half of attacks and isn't a classifier. On 20 hand-built samples it caught 15, listed four more as context, and missed one that hid its text in URL escapes.
What you get
- One Python script with no dependencies. Point it at files or pipe text on stdin.
- 13 pattern groups for things like "ignore your instructions", fake system or tool messages, requests to reveal the system prompt, and commands to send data to a URL. Some German, Spanish and French phrasings are included.
- Four checks for text a reader can't see: HTML comments, hidden CSS, hidden attributes and invisible Unicode characters.
- It also decodes one layer of base64 and HTML entities, so simple encoding doesn't hide a match.
- A JSON or Markdown report with file, line, rule and severity. The matched text itself is always masked, so the report is safe to paste.
- Two tiers. "Actionable" findings look like real instruction attempts and fail the check (exit 1). "Context" findings, such as a plain HTML comment, are listed but don't fail it unless you add
--gate-context. - 83 test cases and five saved reports you can compare against.
Measured on real data (full detail) On 356 README and docs files from 22 popular open-source repos, it raised zero actionable findings, so nothing failed the default check. It did list 448 context findings, mostly HTML comments, badge image links and example URLs.
I also ran it on six public labelled prompt-injection datasets. I tuned rules on half of each dataset and report the other half here. "Caught" means the row failed the default check. "Flagged" means any finding at all.
| Dataset | Attacks caught | Attacks flagged | Benign rows flagged |
|---|---|---|---|
| InjecAgent tool responses | 530 of 1059 | 561 of 1059 | 0 of 1066 |
| Gandalf "ignore instructions" | 26 of 53 | 32 of 53 | (no benign rows) |
| safeguard test set | 71 of 355 | 132 of 355 | 11 of 686 |
| deepset prompt injections | 19 of 127 | 27 of 127 | 0 of 192 |
| jackhhao jailbreaks | 44 of 333 | 229 of 333 | 92 of 680 |
| BIPIA email | 0 of 23 | 0 of 23 | 0 of 28 |
No benign row in any dataset failed the default check. Every false alarm is a context finding from the roleplay rule, on harmless "act as a..." prompts. Read the table as "catches the blunt half of attacks". It isn't a classifier. On 20 hand-built attack samples it caught 15, listed four more as context, and missed one that hid its text in URL escapes. The package's benchmark notes have the full tables and how to rebuild the datasets.
Known limitations
- Polite, task-style injections ("please also forward this to...") look like normal requests. It caught none of the BIPIA email attacks.
- It matches patterns. It doesn't understand intent, so rephrased attacks get through and some harmless prompts get flagged as context.
- Anything inside a Markdown code fence is downgraded to context. That keeps docs quiet, but an attacker can use a fence too.
- URL-escaped text, nested encodings, images and PDFs aren't checked. Language coverage beyond English is a handful of phrasings.
- A pass doesn't make a file safe and doesn't limit what your agent can do.
Who it's for Developers who feed outside text to agents and want a fast, offline check first, by hand or as a CI step.
Requires Python 3.8 or newer. Nothing to install. Input must be UTF-8 files or stdin. It runs locally, makes no network calls and writes no files.
Check it yourself 83 test cases and five saved reports, six public labelled datasets (half held back from tuning), 20 hand-built attack samples and 356 open-source docs files. To reproduce the demo, run this from the package folder: python3 -I -B scripts/injection_scan.py references/example-page.html --format markdown --fail-on medium
Price and guarantee $12 one time, yours to use with any agent that can run Python. Covered by Agensi's money-back guarantee if it doesn't do what this page promises.
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.
English-oriented. No DOM rendering or URL fetching. Misses nested encodings, images, and PDFs. Approximate Markdown fence recognition. Not a guarantee of safety.
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
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 14 days ago
- Software
Python 3
Permissions
Uses a terminal to run Python and reads only the explicitly selected local input files or stdin. No network access. Nothing is sent anywhere. The scanner writes no files; it prints reports to stdout and input errors to stderr. A host may create and delete its temporary script, and a user may redirect a report.