Discussion about this post

User's avatar
Cyril Simonnet's avatar

Agents hitting 1,200 sites in one campaign is the offensive half of the story, and logging agent incidents is the first sign defenders are treating agents as actors that need their own audit trail. The next step follows naturally. The autonomy that lets an attacker's agent scale also makes our own agents a target, because anything they read can become an instruction. A webpage or a ticket comment is enough. I wrote about how agents end up taking orders from strangers, and why the fix starts with treating every input as untrusted and giving each agent the narrowest permissions it can work with. Incident logs like OpenAI's are where that discipline becomes measurable.

https://cyrilsimonnet.substack.com/p/your-ai-agent-takes-orders-from-strangers?utm_source=substor&utm_medium=substack&utm_campaign=comment

Ready for more?