3 ms·

How so? From what I understand, all requests have to be proxied through OneCLI so that agent can't see your keys.

• garettmd 6mo ago
  It used to be that people didn't read the article, now they don't even read the headline.
• rvz 6mo ago
  OneCLI assumes that the proxy is fully trusted by the agent and it still has authorized access to your accounts. What happens when the agent environment is breached? All you need is the fake key + URL of the proxy and that maps to your real keys and you can make authorized requests outside of the agent. The real keys don't have to be leaked, just the fake ones have to map to the real one; so unless they are rotated, then this is a problem.
  • brumbelow 6mo ago
    Exactly. I appreciate the considerations they have already taken, this is definitely a problem that needs to be addressed as agentic AI continues its warpath. However, this feels to me like widening the attack surface rather than tightening security. I'm going to dig in to this over the next few weeks. Hopefully I prove myself wrong