Audited on 500+ open-source agents · 20+ frameworks · Open source
The security scanner for AI agents
Scan Microsoft, Google ADK, Python, LangChain, CrewAI, MCP servers, Skills and 20+ frameworks for agent-specific vulnerabilities — in 60 seconds.
Free forever for up to 5 scans/month. CLI + MCP server open source (Apache 2.0).
Prompt Injection
agent/chains.py:42Untrusted user input flows directly into the LLM prompt. An attacker can inject instructions to override the agent or exfiltrate data.
prompt = f"Answer this query: {user_input}"
response = llm.invoke(prompt)prompt = TEMPLATE.format(query=sanitize(user_input)) response = llm.invoke(prompt)
Built for the modern AI stack
One scanner. Every framework.
What Inkog finds in your agent logic
Three commands. Full security audit.
Install
Scan
Ship with confidence
Get severity-ranked findings, compliance mapping, and remediation guidance.
Or use our MCP server in
Claude &
Cursor
Build secure AI agents with Claude, Cursor, and Claude Code
Connect the Inkog MCP server and ask your AI assistant to scan, explain, and fix agent security issues — without ever leaving the conversation.
Scan during development
"Scan this agent for security issues." Findings come back in the same chat.
Explain & fix in-flow
"Explain this finding and apply the fix." No tab switching, no CLI.
Verify governance
"Does my AGENTS.md match the code?" Only Inkog answers this.
Every finding mapped to compliance
Inkog doesn't just find vulnerabilities — it tells you which regulation they violate and what to fix.
State of AI Agent Security 2026
The largest security analysis of the AI agent ecosystem. 500+ open-source projects scanned. Free download.
Download the reportAgent Capability Surface
One score, three layers, every gap mapped to a regulation. The first inventory that tells you what your agents can do, what your AGENTS.md says they should do, and where the controls are missing.
Every tool, MCP server, delegation, memory access, and credential the agent can reach. Extracted by the Universal IR across 15 frameworks.
Every line of AGENTS.md, parsed across YAML front matter, markdown sections, and inline annotations into typed declarations.
Every control wired in code: human approval, authorization, audit log, rate limit, cycle guard, sanitizer. Indexed against the capability it protects.
What a scan returns
Paste a GitHub URL or upload a zip. No install, no config. Core results in about 60 seconds; deep analysis typically takes about 10 minutes and can run up to 30. Here's what comes back.
Agent tool executes code via eval/exec where input can be influenced by LLM output or prompt injection.
pickle.load() on a file the agent fetched. A crafted pickle runs arbitrary code when the vector store is loaded.
process.py:73The same FAISS store is unpickled again at request time, so the exposure repeats on every server start.
server.py:15Latest from Labs
Security research, vulnerability disclosures, and technical deep-dives from the Inkog team.
Start scanning in 60 seconds
Free · No setup required · Core results in about 60 seconds
15 min · Live Deep Scan on your code · Walkthrough of the findings