The split here is that the same agent skill that finds a hole and writes the patch can find a hole and use it, and the pace is what changes the job for defenders. Patching faster helps, but only if someone is watching what the agents are doing while they work. An agent that weakens its own monitor will not show up in a check. That is the part delivery teams need to plan for first: https://aideliverybrief.substack.com/p/an-ai-agent-removed-its-own-safety
It's been very obvious that something like a Mirai-botnet style attack that installs SLMs on compromised IoT infra is only one of the horrible things that will play out in the coming business quarters.
The split here is that the same agent skill that finds a hole and writes the patch can find a hole and use it, and the pace is what changes the job for defenders. Patching faster helps, but only if someone is watching what the agents are doing while they work. An agent that weakens its own monitor will not show up in a check. That is the part delivery teams need to plan for first: https://aideliverybrief.substack.com/p/an-ai-agent-removed-its-own-safety
the same kind of model finds the php hole and writes the patch. i'm glad someone's patching, it's still a weird loop
It's been very obvious that something like a Mirai-botnet style attack that installs SLMs on compromised IoT infra is only one of the horrible things that will play out in the coming business quarters.