Cerberus watches what your AI agent's tools do — the private data they read, the untrusted content they ingest, and the outbound calls they make — and blocks a prompt-injection exfiltration before the outbound tool runs. It works at the tool-call level, so it's robust to prompt wording and model changes.

Cerberus Core is now live — runtime security for AI agents that actually works. Most agents are one clever prompt injection away from leaking your data. The classic attack uses the Lethal Trifecta: Privileged data access Injected instructions from untrusted content Outbound exfiltration Cerberus watches the tools themselves at runtime, correlates all three, and blocks the outbound call before a single byte leaves — completely independent of prompt wording or model changes. Open core (MIT) npm / PyPI ready LangChain, Vercel AI SDK, OpenAI Agents, LangGraph, CrewAI, etc. Try the demos in < 60 seconds Repo: https://github.com/Odingard/cerb... Product / Enterprise: https://sixsenseenterprise.com If you build agents that touch real data + tools, this is your new guardian. Drop a comment if you're integrating it — happy to help with setup! #AISecurity #AgentSecurity #PromptInjection #LLM #OpenSource

Took it for a spin with a sketchy agent setup and it actually caught an exfiltration attempt when the model tried to phone home after reading a poisoned doc. The tool-level interception feels much more reliable than the prompt-parsing stuff I have tried before.

@esila2293619488 thank you very much! Please run it hard! This is just the core version as the enterprise paid version has more! Thank you again

Cerberus

Cerberus watches what your AI agent's tools do

Cerberus watches what your AI agent's tools do

Forum Threads

Feedback and Challenges

Forum Threads

Feedback and Challenges