AI Agents Exploit Vulnerabilities in Hours, Compromising 395 Organizations Across 48 Countries

AI Agents Turn Vulnerability Exploitation Into Hours-Long Operations

On August 31, a single operator deployed a few hundred AI agents targeting a print-management server used globally by schools and offices. The AI agents built a test lab, wrote exploits for two bugs disclosed just four days earlier, pulled a target list from an internet scanning service, and launched their campaign1

1

. From an empty workspace to the first break-in took under four hours. When the campaign launched the next morning, 11 organizations fell in 26 seconds. By the time GreyNoise finished counting, 395 organizations in 48 countries had been compromised, with half of them schools1

1

. The model executing the work was a cheap, open-weight one, not a frontier model from a US lab, and the AI agents even ignored a list of 28 countries they were instructed to avoid.

PHP2Shell Attack Exposes Critical Gaps in Cybersecurity Defense

Seven weeks later, two critical bugs in PHP, the programming language used by 70% of websites, leaked ahead of their patch1

1

. Exploit attempts were hitting Amazon's decoy servers four days before disclosure and 12 hours before the fix, with working shells following within three days. CJ Moses, Amazon's CISO, wrote that AI "makes it cheaper to turn a patch diff or advisory into a working exploit" and that "a 7-day or 30-day SLA assumes a delay that no longer exists"1

1

.

The PHP2Shell vulnerabilities, which let an attacker take over a web server, had lurked in the language since at least 2012 until a researcher working with an OpenAI model found them2

2

. PHP's maintainers released a patch on October 21, but attacks were already under way. Qualys counted 167 million PHP installations on its customers' systems. Twelve hours after disclosure, about one in a thousand had been updated2

2

.

The Collapse of the Exploit Bottleneck

For twenty years, the bottleneck in offensive cyber was the expertise and time required to turn a vulnerability into a working exploit. In 2025 and 2026, that work became cheap and quick, minutes or hours instead of days or months1

1

. RAND's 2017 study of more than 200 zero-days found a median of 22 days for a professional to build a working exploit, with brokers paying thousands to millions of dollars each1

1

.

Mandiant's data shows the average time from disclosure to attack falling from 63 days in 2018-19 to 44 in 2020-21, 32 in 2021-22, and five in 2023. By 2024, attacks came an average of one day before disclosure. By 2025, seven days before1

1

. AI can now derive from a patch what vulnerability it fixes and how to exploit it, work that once required intelligence services or elite hacking groups2

2

.

AI-Driven Cyberattacks Reshape Threat Landscape

AI is cutting the cost of every step in the attack chain, not just the exploit. Scouting a network, adapting an exploit to an unusual setup, moving inside, getting paid: each step needed a person, and a person's days only made sense against a target that could pay them back1

1

. Two of the largest incidents this fall needed no new vulnerability at all: the supermarket refrigeration attacks hit controllers that no longer get updates and whose weaknesses had been public for months, and the Binance theft went through one engineer's laptop1

1

.

On October 27, Anthropic said its Claude model had written an exploit that seizes control of some Linux servers through a vulnerability in how they handle network connections. Linux's developers fixed it on October 5. As of November 2, Ubuntu, one of the most widely used versions of Linux for servers, still had not shipped the fix2

2

. Jason Clinton, Anthropic's deputy chief information security officer, wrote that things "are going to get a lot worse before they get better," and urged security leaders to "go to your boards and ask for help right now"2

2

.

Media Coverage Gap Leaves Public Unaware of Cyber Risks

In the eight days after PHP2Shell became public on October 20, CNN, Fox News, NPR and USA Today published 5,610 stories, and only nine mentioned it or the week's other major vulnerabilities. Two put them in a headline2

2

. A record 5,461 vulnerabilities were published in the week of October 12 alone, more than 15 times the weekly average in 20202

2

.

Experts argue the public needs the equivalent of a hurricane watch. CISA, the federal Cybersecurity and Infrastructure Security Agency, should issue a public watch when a critical vulnerability turns up in widely used software, and a warning when attacks begin2

2

. Unpatched vulnerabilities are where many damaging attacks begin. In March 2017, the Apache Software Foundation patched a vulnerability in Struts. Two months later, attackers used it to break into Equifax, which had not installed the patch, and went on to steal personal data on 147 million Americans2

2

.

What Organizations Should Watch For

Patching did not get faster, and for a large share of the world's systems, especially critical ones, it can't1

1

. Defenders have not kept up. The gap between patch release and deployment can last weeks, creating windows of exposure that AI-driven cyberattacks now exploit with unprecedented speed. Watch for outbreaks similar to the 2003 worm era and the 2017 WannaCry outbreak, landing hardest on systems that cannot be updated quickly, and on the few targets where a single break-in is worth billions1

1

. The discovery and exploitation of vulnerabilities now happens faster than organizations can respond, fundamentally changing cybersecurity threats and requiring boards and lawmakers to reassess how they allocate resources to cyber defense.