👀 OpenAI just opened agent-building capabilities to every one of your employees with ChatGPT access, dramatically increasing the AI attack surface. 👀 OpenAI AgentKit flow agents typically operate with broad permissions across multiple systems. When combined with common misconfigurations the blast radius becomes massive: ‼️ Missing approval gates for destructive operations ‼️ Un-sanitized data flows to external MCP servers ‼️ Prompt injection vulnerabilities in workflow logic ‼️ Data access patterns that bypass user permissions AI agents configured and deployed by business users using NCLC platforms can expose attack vectors associated with business critical functions such as expense approval logic, customer support workflows, and data access policies, circumventing traditional application or data security guardrails. The question isn't whether agent sprawl is coming. It's whether you'll have visibility and control when it does. Our latest blog post from Gal Moyal breaks down the hidden attack surfaces of OpenAI AgentKit, with proven and realistic recommendations for embedded agent security. Get the link in the top comment below.
Maor Volokh
1w
Great insights from Gal Moyal 👑
Gal Moyal
1w
If you don't design your workflow carefully, it might just expose your organization to risk. See common misconfigurations here!
Read the blog post from Gal Moyal here to learn how to secure OpenAI AgentKit agents: https://noma.security/blog/openai-agentkit-just-democratized-agent-building-and-multiplied-your-attack-surface/