AI has uncovered a flaw nearly 30 years old that no one had noticed until now

Sometimes computer code hides secrets that elude human eyes for decades. It’s exactly what happened with Squid, one of the most widely used open-source caching proxies in the world.

A critical vulnerability, dormant since the late ’90s, has exposed for as long as 29 years users’ unencrypted HTTP requests, potentially allowing leakage of extremely sensitive data such as passwords and session tokens.

A break in this long chain of insecurity was provided by a researcher, supported by Anthropic’s artificial intelligence Mythos Preview, who identified the anomaly and allowed developers to take steps to remedy it.

Mythos finds what has eluded humans for nearly 30 years

It all began in an unusual scenario: during a routine commercial flight. The researcher Lam Jun Rong, a specialist at the security company Calif.io, was trying to connect to the airborne network when he stumbled upon an installation of Squid (a software widely adopted by large corporations, schools, and Internet service providers to filter, monitor, and store traffic across the network) dating back almost a decade.

Intrigued by the discovery, he launched a thorough analysis of the proxy.

Working closely with the AI agent Mythos Preview, the researcher was able to isolate a structural flaw he himself dubbed “Squidbleed“.

The vulnerability, officially tracked under the identifier CVE-2026-47729, was promptly reported in April to the maintainers of the original project. The development team eventually released the fix patch on June 8, integrating it into the version 7.6 of the software.

The NetWare Ghost and the bug in the code

The most fascinating aspect of this story lies in the bug’s very origin. The flaw was nested within the parser dedicated to reading Squid’s FTP directories and was introduced inadvertently through the commit identified by the string bb97dd37a, compiled in the distant 1997.

In that era, the source code was modified to ensure compatibility with NetWare servers, a network operating system that enjoyed enormous popularity in the corporate market before the eventual dominance of Windows and Linux.

These old servers returned data by adding extra blank spaces between the modification timestamp and the actual filename, differing from normal formatting standards. To bypass the obstacle, a loop was inserted specifically to ignore those excess spaces.

However, the joint human-artificial analysis revealed a microscopic yet deadly detail: if an FTP server controlled by an attacker maliciously omits providing the file name, the system points directly to the null termination character.

The search function used at that point treats that character as if it were part of the text string, causing an infinite loop that overruns the buffer and returns to the hostile server fragments of internal memory loaded with private information.

Attack surface and risk mitigation

For such intrusions to succeed, the network infrastructure must meet two specific conditions. First, the proxy must have the ability to inspect traffic, so it must process plaintext requests or be in an environment where encryption is terminated locally. Second, the system must allow outbound communication via TCP port 21, reserved for the FTP protocol.

This file transfer standard is today a true technological relic, to the extent that Chromium-based browsers have discontinued its basic support for many years.

The software fix introduced with the latest Squid release consists of a simple preventive check before reading the text.

Cybersecurity experts strongly recommend system administrators not only to install the new release immediately, but also to isolate the root of the problem by disabling FTP access entirely within proxy configurations.

Considering that modern organizations generate virtually no legitimate traffic to that protocol, its complete deactivation immediately closes the entire attack surface without causing any impact or service disruption for end users.