Organizations “can’t patch their way out of AI” and should instead focus on mitigation post exploitation, according to Dick O’Brien, principal intelligence analyst at Symantec.

Speaking at the IRISSCON 2026 conference in Dublin, Ireland, on November 11, he highlighted the rapid closure of the patch window in the past year due to attackers’ use of AI.

This includes AI tools to discover zero day flaws and the ability of frontier models to chain together low severity vulnerabilities together to find an exploit path in a way that humans would not be able to do.

The Zero Day Clock, a live dashboard that tracks the collapse of the time-to-exploit (TTE) window from CVE disclosure to confirmed exploitation in the wild, has estimated that the average time to exploitation following public disclosure of a vulnerability fell from one week in 2025 to one day in 2026. The initiative estimates that in 2027, the average time will be a minute.

“Once you get to a minute, what can you do realistically? We need to start thinking beyond patching to address vulnerabilities,” O’Brien said.

This shift was demonstrated by the speed at which attackers exploited the recent PHP2Shell vulnerability following its disclosure in October 2026. O’Brien noted that the first reporting exploitation attempts occurred just 19 minutes after the bug was made public.

Amazon also reported that it detected an attempted exploitation attempt of PHP2Shell four days before the vulnerabilities were published.

Agentic AI Orchestration of Attacks

Post-initial access, O’Brien expects threat actors to move to an era of “end-to-end agentic orchestration of attacks” in the next year, shifting away from using AI as an assistant for various tasks towards AI agents undertaking attacks independently.

This will make it significantly harder for defenders to detect and respond to threats in their organization’s environments.

There have been a few examples of this approach taking place so far. One of these was an attack on the Dutch Institute of Vulnerability Disclosure (DIVD) in September 2026, where AI agents executed the post-compromise tasks independently, operating at machine speed.

According to an analysis by Symantec of the DIVD’s assessment report, it took 19 minutes for the attack to be detected, but 19 hours before the response began, and a further nine hours to contain the threat.

“This is what we’re dealing with. It’s an excess of 24 hours and you know you’re not going to be able to compete with something operating at machine speed,” O’Brien commented.

Adopt an Assume Exploit Mindset

In this landscape, organizations should assume exploitation is inevitable and prepare for machine-speed post-compromise activity, where attackers can move from initial access to impact in minutes rather than days.

“Stop measuring your mean time to patch and start measuring your mean time to mitigate. Assume an exploit is going to happen and the plan your preparedness for that,” O’Brien advised.

He set out a number of mitigation tasks security teams can immediately focus on:

• Behavior blocking
• Default-deny for what runs
• Least privilege for every agent and token
• Inspect data before it leaves the device
• Bind every agent action to a named human

“These are all things you can start doing right now before you become aware of a vulnerability,” O’Brien added.

Longer-term, organizations must integrate detection and protection technologies on multiple different control points, which are able to identify behavioral anomalies of rogue or malicious AI agents.