Classic cybersecurity metrics, developed over decades of practice, are no longer consistent with reality. According to a CrowdStrike report published in November 2026, attackers are increasingly using autonomous AI agents that execute up to 1,500 commands per hour — dozens of times faster than a human. As a result, standard indicators such as breakout time or the 1-10-60 rule are losing their meaning, forcing defenders to rethink the very logic of assessing response effectiveness.

Campaign Against Argentine Hospitals: 29 Seconds to Change Tactics

CrowdStrike specialists identified four previously undescribed attacker groups that used the Claude Code agent platform to conduct attacks between May and August 2026. The most striking episode occurred during a campaign against medical organizations in Argentina: after blocking the first infiltration attempt, an agent reconfigured its tactics in 29 seconds, and 17 minutes later it already held five domain controller credential hashes. In total, at least 15 medical institutions were compromised, with credentials and patient medical records stolen from at least one of them. The attack architecture was two-tiered: a Supervisor agent distributed tasks, and an Executor executed them. These systems were discovered thanks to publicly accessible directories containing agent platform files left in the victims' infrastructure.

Breakout Time Collapse: From Two Hours to 27 Seconds

The breakout time indicator — the time between initial infiltration and an attacker moving laterally through the network — was introduced by CrowdStrike in 2018. Back then, the average was 1 hour and 58 minutes. In the latest report, the average figure has dropped to 29 minutes, and in the fastest recorded case, the internal network attack took 27 seconds. Company CEO George Kurtz stated outright that this metric no longer reflects reality, as the boundary between the speed of human and machine actions has been practically erased. In the second half of 2026, CrowdStrike registered three times more intrusions involving AI agents than in the first. The 1-10-60 rule — detect a threat in a minute, investigate in ten, and remediate in sixty — has also stopped matching the pace of attacks, and the company is currently discussing its replacement.

Access Transfer and Data Theft Speed: Numbers That Change the Picture

The problem is not limited to movement within a single network. The median time to transfer already obtained access from one attacker group to another has shrunk from over eight hours in 2022 to 22 seconds in 2025. Palo Alto Networks notes a similar trend: in the fastest investigated cases, attackers needed 72 minutes to transition from initial access to data theft, four times faster than a year earlier. Amazon, in turn, reported an attack that an AI agent carried out in 12 minutes and 42 seconds: the AWS honeypot system registered 94 events without a single syntax error — a sign that the actions were generated by a machine, not a human.

The Bottleneck of Defense: Not Detection, But Follow-up Actions

Experts suggest abandoning a single overall attack speed metric and measuring the duration of each defense stage separately: detection, investigation, decision-making, localization, and remediation. The case of the Dutch Institute for Vulnerability Disclosure (DIVD) is telling: in September, attackers used two zero-day vulnerabilities to gain access to a ticketing system, after which AI agents carried out further actions. It took the DIVD team 19 minutes to detect the intrusion, 19 hours to prepare the first response, and another 9 hours to localize the attack. The example clearly demonstrates that the bottleneck is not detection speed, but the subsequent organizational and technical actions of defenders.

The Gap Between Exploitation and Vulnerability Patching

The second critical indicator is the time from vulnerability disclosure to its remediation. According to Verizon data, the median time to fully patch an already exploited vulnerability in 2026 was 43 days, compared to 32 days a year earlier. Another study found that the average time from public disclosure of a vulnerability to the first exploitation attempt dropped from 115 days in 2024 to 44 days in 2025. Comparing these two datasets reveals a dangerous gap: attackers are gaining the ability to use known issues faster and faster, while companies take weeks to install patches. In June 2026, the US Cybersecurity and Infrastructure Security Agency (CISA) revised patching deadlines for federal agencies: for the most dangerous vulnerabilities, including those where exploitation can be automated, a three-day deadline was set.

Against the backdrop of these data, the issue of managing the access of AI agents themselves to corporate systems becomes operational rather than theoretical: if an attacker's agent can traverse from the network perimeter to a domain controller in 27 seconds, then an agent integrated into the defender's infrastructure will require no less stringent control of permissions, logging, and isolation. The transition from a single speed metric to a multi-stage assessment of every link in the response chain essentially means reassembling the entire cybersecurity methodology to conditions where the opponent acts orders of magnitude faster than a human.