More screenshots
Works with the AI Agents you already use
Agent Security Review Pack — Score Your Agent Build
Reads your agent repo and scores the surfaces an injected instruction can reach: prompt assembly, tool permissions, secrets, output handling, run bounds and egress, mapped to the OWASP Agentic Top 10.
$14.99
Agent Security Review Pack — Score Your Agent Build
Example session with this skill installed
python3 scripts/security_review.py . --report security-review.md
Repository
app.py prompt built with customer email
tools.json send_email, filesystem root "/"
render.py innerHTML = response
agent.py while True: ...
logs.py logger.info(email)
- Read your context and instructions
- Compiled the agent security review
agent-security-review agent-app (6 files)
score 19/100 grade F findings 7 high 4 medium 3
OWASP ASI01, ASI02, ASI08, LLM02, LLM03, LLM05
HIGH S1 Untrusted content is merged into the system prompt app.py:5 [ASI01]
HIGH S2 A write-capable tool has no approval gate tools.json:2 [ASI02]
HIGH S4 Model output is rendered without escaping render.py:3 [LLM05]
HIGH S6 A tool grants filesystem access it does not need tools.json:3 [LLM03]
MEDIUM S5 The agent loop has no step cap agent.py [ASI08]
MEDIUM S7 Personal data is written to logs logs.py:5 [LLM02]
FAIL findings at or above 'high'
report security-review.md
What to do next
- Move the customer email out of the system prompt and pass it as data.
- Require an approval bound to the actor and arguments before send_email.
- Render the answer as text; sanitise only with a maintained allowlist.
- Scope filesystem_read to the reports directory.
- Cap steps and spend in orchestration, then re-run for a new score.
Connects securely to your tools. The creator never sees your data.
About this skill
A checklist you read is not a review. This walks the repository, collects the surfaces that decide what an injected instruction can actually do, and returns a score with file-level evidence and the OWASP Agentic category each finding maps to.
- S1 prompt assembly: untrusted text merged into the instruction channel (ASI01 goal hijack)
- S2 tool permissions: a write-capable tool with no approval gate (ASI02 tool misuse)
- S3 secrets in prompts or config, because the key's reach becomes the agent's reach (LLM07)
- S4 output handling: model output rendered unescaped into the DOM or a template (LLM05)
- S5 run bounds: an agent loop with no step or tool-call cap (ASI08)
- S6 shell or root filesystem access a tool does not need (LLM03 excessive agency)
- S7 personal data written to logs, and S8 outbound URLs fetched without an allowlist
Start at 100, deduct by severity, and get a grade plus deductions grouped by domain, so the first fix is obvious. Every finding carries the exact line, the category, a confidence level and the control that closes it.
The pack ships two builds: a deliberately weak agent that scores 19/100 (F) with four high findings, and the same app with the controls in place at 100/100 — so you can see what a reviewed surface looks like before you run it on yours.
Good to know: this is a static review of prompts, configuration and code patterns. It does not prove exploitability and does not replace a penetration test. It does not review model behaviour, runtime policy enforcement or infrastructure, and it cannot see services whose code is not in the directory. Medium-confidence findings are heuristics and should be read before acting. The score is a prioritisation aid, not a certification.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
Static analysis only. Does not prove exploitability or model-side safety. Requires Python 3.10+. Limited to code/configs present in the scanned directory.
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 1 day ago
- Software
Python 3.10+An agent repository to review