It was done in 3 hours and 49 minutes. That was recorded as the time Cipher, an AI agent from security startup Causal Security, took to find all the vulnerabilities in software that controls power grids and accelerators. The US Cybersecurity and Infrastructure Security Agency (CISA) warned of a remote code execution risk in the software, and version 7.0.11, which contains the patch, was released on the same day as the advisory.
A job started at 9 a.m. would finish at 12:49 p.m. An AI agent is a program that, given a goal, carries out multiple steps of work on its own, and remote code execution is a flaw that lets a distant attacker run their own commands on someone else's system. If a search ends in less than half a day, the list of flaws could grow faster than people can review it.
Until it is disclosed when the clock was started and stopped and how much people assisted, 3 hours 49 minutes should be read as a reference figure.
On November 2, a zero-day incident involving the web development language PHP was reported. A zero-day is a vulnerability that comes to light before a patch is available to fix it, and defenders must hold out with no fix to download. There is no basis for grouping it with the power grid warning as the same incident.
Automated intrusion attacks using AI agents were also reported as an example of AI-driven security issues. These are cases in which both the side finding vulnerabilities and the side breaking through them used AI agents.
The achievement shown in the Cipher case extends as far as finding flaws. Even when a patch such as 7.0.11 is released, the risk does not fall until it is installed where the software is actually in use. If only discovery speeds up, human approval and deployment procedures will inevitably stand out as the slowest stage.
Oversight is also a challenge. A report on November 1 said a "personal AI agent" had run for two months on a security network of the US Department of Defense (the Pentagon), and it has not been determined whether it was approved or whether it led to a breach. If a tool exists that can complete a 3-hour-49-minute job, two months is an extremely long time.
A report also came out the same day that financial authorities are preparing a hacking defense strategy against AI agent threats. When the strategy is released, the thing to check is whether it addresses how to shorten the time from discovering a flaw to applying a patch and how to oversee agents running inside the financial sector.
The workday of security staff could also change. If they arrive at 9 a.m. and run an agent, a list of flaws appears on the monitor around lunchtime, and the afternoon is filled with sorting out and approving which flaws to fix most urgently. The more that finding flaws is left to machines, the more of people's working hours will go to fixes and approvals.
