Insights / Real-time cybersecurity
Real-time cybersecurity

An agent swarm took 14 minutes. Spotting it took 4.5 hours.

A security executive says a swarm of more than 30 agents got past a company’s two-factor login, stole data and encrypted its files in 14 minutes, and the company needed about 4.5 hours to notice. Real time can’t mean a faster meeting. It means decisions made in advance.

Ausper Technology · Nov 20, 2026 · Insight · Real-time cybersecurity · 7 min read

The short version
  • On CNN, David Kennedy described an agent swarm that got past a company’s 2FA, stole data and encrypted its files in 14 minutes. Noticing took about 4.5 hours.
  • It is one unverified account, but the data agrees: a 29-minute average breakout, 72 minutes to stolen data for the fastest quarter of intrusions, a 14-day median dwell time.
  • Real time means deciding in advance: see in seconds, pre-approve narrow, reversible actions, and use phishing-resistant MFA, as OMB M-22-09 already requires.

On CNN’s Erin Burnett OutFront on November 19, David Kennedy, founder and chief executive of TrustedSec, whom CNN introduced as a former NSA hacker, described a breach he is working for a client he can’t yet name. The attackers launched “what we call an agent swarm”: “over 30 different types of agents probing their infrastructure at the exact same time”. They logged in with valid usernames and passwords, found a flaw in how the company’s application handled two-factor authentication, took a substantial amount of data, and then set off ransomware that encrypted the entire file system. “And this all took 14 minutes, 14 minutes. I’ve never seen anything that fast before” (CNN transcript).

Then the line that matters most: it “took about 4-1/2 hours for them to actually identify that it occurred”. Fourteen minutes to take the network, about 270 to notice. This is one responder’s account, not independently verified (IBTimes UK says the same), and the company and the technical details are not public. We treat it as a marker, not a measurement. The measurements point the same way.

Fig. 1
Noticing the attack took about 19 times longer than the attack.
270 minabout 4.5 hours for the company to identify it
14 minthe whole attack, from probing to encrypted files
19×longer to see it than to do it
Source: David Kennedy on CNN’s Erin Burnett OutFront, Nov. 19, 2026 (rush transcript). Ausper calculation.
Data
StepValueWhat it is
1270 minabout 4.5 hours for the company to identify it
214 minthe whole attack, from probing to encrypted files
319×longer to see it than to do it

The attacker’s clock now runs in minutes

CrowdStrike’s 2026 Global Threat Report puts the average eCrime breakout time, from first access to moving across the network, at 29 minutes in 2025, with the fastest at 27 seconds; in one intrusion, data theft began within four minutes. Unit 42 found the fastest quarter of intrusions reached data exfiltration in 72 minutes in 2025, down from 285 in 2024, and 22 percent of incidents got there in under an hour. The work is also being handed to machines: Anthropic reported in November 2025 that a state-sponsored group used AI to perform 80 to 90 percent of an espionage campaign against roughly 30 targets, often several requests a second. Its September 2026 report says that operating model has since spread to every class of actor it investigated.

Fig. 2
Attackers move in minutes. The median intruder goes unseen for 14 days.
14 minProbing to encrypted files, in the breach Kennedy described on CNN.CNN
29 minAverage eCrime breakout time in 2025, first access to moving across the network.CrowdStrike
27 secThe fastest breakout CrowdStrike observed in 2025.CrowdStrike
14 daysGlobal median dwell time in 2025: how long an intruder stays before anyone notices.Mandiant
Sources: CNN Erin Burnett OutFront transcript, Nov. 19, 2026; CrowdStrike 2026 Global Threat Report; Mandiant M-Trends 2026.
Data
NumberWhat it isSource
14 minProbing to encrypted files, in the breach Kennedy described on CNN.CNN
29 minAverage eCrime breakout time in 2025, first access to moving across the network.CrowdStrike
27 secThe fastest breakout CrowdStrike observed in 2025.CrowdStrike
14 daysGlobal median dwell time in 2025: how long an intruder stays before anyone notices.Mandiant

Detection hasn’t kept pace. Mandiant’s M-Trends 2026 puts the global median dwell time, how long an intruder stays before anyone notices, at 14 days. Against that, the 4.5 hours in Kennedy’s account is fast. It still came 4 hours and 16 minutes after the movie was over.

Fig. 3
From break-in to stolen data: 285 minutes in 2024, 72 in 2025, 14 in this account
Fastest quarter of intrusions, 2024Unit 42 incident response cases
285 min
Fastest quarter of intrusions, 2025Unit 42 incident response cases
72 min
Unit 42’s AI-assisted attack testA controlled experiment, not a real case
25 min
The breach Kennedy describedData taken and files encrypted, his account on CNN
14 min
Sources: Unit 42 Global Incident Response Report 2026 (2024 and 2025; the 25-minute test was first reported in 2025); CNN transcript, Nov. 19, 2026.
Data
CaseMinutesNote
Fastest quarter of intrusions, 2024285 minUnit 42 incident response cases
Fastest quarter of intrusions, 202572 minUnit 42 incident response cases
Unit 42’s AI-assisted attack test25 minA controlled experiment, not a real case
The breach Kennedy described14 minData taken and files encrypted, his account on CNN

2FA is not the finish line

In this case the attackers already had working passwords, and the second factor failed because of a flaw in how the application implemented it, Kennedy said. He gave no more detail, and we won’t guess. The general pattern is well measured. Microsoft’s Digital Defense Report 2026 found adversary-in-the-middle phishing in 44.6 percent of identified phishing techniques: a fake sign-in page relays the code or push approval to the real site and keeps the session cookie that comes back. Unit 42 found identity weaknesses in nearly 90 percent of its investigations.

The federal answer has been on paper since 2022. OMB M-22-09 says: “For agency staff, contractors, and partners, phishing-resistant MFA is required”, and tells agencies to drop methods that fail to resist phishing, including SMS, voice calls, one-time codes and push notifications. Phishing-resistant means PIV or CAC smart cards and FIDO2 security keys or passkeys, which CISA calls the gold standard. Two more lessons come straight from this case: test how each application enforces the second factor, not only which factor it uses, and be able to revoke every session and token for an account in one step, because a stolen session skips MFA entirely.

Real time: the movie version and the real one

In the movies, real-time security is a wall of glowing screens and someone shouting “they’re in”. The screens are the easy part. A 14-minute attack is decided by what happened before it started, because no bridge call, approval chain or change board can meet, decide and act inside that window. Real time means three things settled in advance:

  • See in seconds. Sign-ins, endpoints, cloud control planes and AI agents report continuously, and someone, or something, is watching at 2 a.m. on a holiday. Thirty agents probing at once is a pattern a watching system can flag.
  • Act without asking, inside limits you set. A short list of narrow, reversible moves your tools may make on their own: revoke an account’s sessions and tokens, isolate a host, disable an account, cut an AI agent’s keys. Approve the list now, with the people who would otherwise be woken at 2 a.m. to approve each one.
  • Prove it as it runs. Evidence comes from the systems themselves, so the security team and the authorizing official see the same picture, not last quarter’s document.

The fair objection is that speed breaks things, and this week shows it. The NSA’s classified internal network has been disrupted since November 17 (CNN) and was reported still down on November 19 (NBC News); officials tied it to patching that did not come back up as expected, not to an attack. That is why the automatic list stays narrow and reversible, gets rehearsed like a fire drill, and every action is logged and reviewed. A session revoked by mistake costs someone a sign-in. Fourteen minutes of waiting can cost the file system.

How we work on it

That gap is what our real-time cybersecurity practice is built around. We work it from several directions at once: GRC with live posture and continuous risk scoring (AusperRiskIQ™), security operations and vulnerability response around the clock, and authorization evidence that flows from your pipelines instead of aging in a document. Our Next-Generation GRC prototype (beta) shows the operator’s view we design for: posture captured in real time, and the systems a finding affects known in real time. And our threat globe is the movie screen built on real feeds: it reads CISA’s exploited-vulnerability catalog and advisories every hour and puts them on a live globe.

For a federal program the first moves are usually the same: time the response loop end to end, write down and approve the actions tools may take on their own, and close the gaps in phishing-resistant MFA, starting with administrators and remote access.

Five questions for Monday

  • How many minutes pass between a suspicious sign-in and someone, or something, acting on it? Time it on a real test.
  • Which actions can your tools take without a person? Is that list written down and approved?
  • Do administrators and remote access use phishing-resistant MFA (PIV, CAC or FIDO2), as M-22-09 requires, and has anyone tested how each application enforces it?
  • Can you revoke every session and token for one account, including the keys its AI agents hold, in one step?
  • Who is watching at 2 a.m. on a holiday?

Common questions

Is the 14-minute breach confirmed?
Not independently. It is David Kennedy’s account on CNN of an incident he says he is responding to; the company and the technical details are not public. The argument here rests on measured data: CrowdStrike’s 29-minute average breakout, Unit 42’s 72 minutes to exfiltration for the fastest quarter of intrusions, and Mandiant’s 14-day median dwell time.
How was the two-factor authentication bypassed?
Kennedy said only that the attackers logged in with valid usernames and passwords and found a flaw in how the company’s application handled two-factor authentication. No other detail is public.
What counts as phishing-resistant MFA?
Authentication a fake site can’t relay: PIV or CAC smart cards and FIDO2/WebAuthn security keys or passkeys. OMB M-22-09 requires it for agency staff, contractors and partners, and tells agencies to drop SMS, voice, one-time codes and push notifications.
Won’t automatic containment cause outages?
It can, which is why the actions should be narrow and reversible (revoke sessions, isolate a host), rehearsed, and logged for review. A mistaken session revoke costs a sign-in; waiting costs the 14 minutes.

Sources

Last verified Nov 20, 2026

  1. CNN, Erin Burnett OutFront rush transcript, Nov. 19, 2026 (David Kennedy: an agent swarm, 14 minutes, about 4.5 hours to identify)Accessed Nov 20, 2026.
  2. IBTimes UK, AI attack bypassed two-factor authentication in 14 minutes (account not independently verified), Nov. 20, 2026Accessed Nov 20, 2026.
  3. TrustedSec, David Kennedy, Founder and CEOAccessed Nov 20, 2026.
  4. CrowdStrike, 2026 Global Threat Report press release (29-minute average breakout in 2025; fastest 27 seconds; exfiltration within four minutes in one intrusion), Feb 24, 2026Accessed Nov 20, 2026.
  5. Palo Alto Networks, Unit 42 Global Incident Response Report 2026 press release (72 minutes to exfiltration in the fastest attacks), Feb 17, 2026Report figures: fastest quartile 72 minutes in 2025, 285 in 2024; 22 percent under an hour; median two days; the 25-minute AI-assisted test was first reported in 2025. Accessed Nov 20, 2026.
  6. Palo Alto Networks, Unit 42 Global IR Report blog (identity weaknesses in nearly 90 percent of investigations), Feb 17, 2026Accessed Nov 20, 2026.
  7. Anthropic, Disrupting the first reported AI-orchestrated cyber espionage campaign, Nov 13, 2025Accessed Nov 20, 2026.
  8. Anthropic, Detecting and countering misuse of AI: September 2026Accessed Nov 20, 2026.
  9. Microsoft, Digital Defense Report 2026 (AiTM at 44.6 percent of identified phishing techniques, p. 45)Accessed Nov 20, 2026.
  10. Google Cloud (Mandiant), M-Trends 2026 (global median dwell time 14 days), Mar 23, 2026Accessed Nov 20, 2026.
  11. OMB M-22-09, Moving the U.S. Government Toward Zero Trust Cybersecurity Principles, Jan 26, 2022Accessed Nov 20, 2026.
  12. CISA, Implementing Phishing-Resistant MFA (fact sheet), Oct 2022Accessed Nov 20, 2026.
  13. CNN, Highly classified US national security communications network goes down, sources say, Nov 17, 2026Accessed Nov 20, 2026.
  14. NBC News, Classified NSA network still down, leaving staff working from home, Nov 19, 2026Accessed Nov 20, 2026.

Updates

  • Nov 20, 2026. Published.

Get Insights by email

Thanks for subscribing.

Share on LinkedInFollow AusperSubscribe by RSS

Related reading

Put this to work

Working on this now?

Tell us where your program stands and we'll tell you what we would do next.

Talk to Ausper