Squidbleed: AI Discovers 29-Year-Old Bug That Escaped Decades of Code Reviews

Squidbleed: The Bug That Nobody Wanted to See – Until a AI Took a Look

On June 23, 2026, researchers from Calif.io published a blog post that shook the security community like a meteorite impact. The title: “Squidbleed – a 29-year-old bug in Squid Proxy”. The content: They had found a mistake, programmed by an developer in 1997, using Claude Mythos Preview from Anthropics. The bug had survived three decades, numerous releases, audits, and rewrites. It can be fixed with a single additional check in the code. And it can still leak passwords and API keys from the storage of millions of networks today.

Why is this more than a nerd story? Because Squidbleed exemplifies what 2026 will bring in cybersecurity. AI models find errors in millions of lines of code that humans have overlooked for decades. Anthropics has discovered alone around 23,000 security vulnerabilities in open-source projects with its Project Glasswing. The Five Eyes security services now say: “Not years, but months” until AI fundamentally changes the cyber world. Squidbleed is a good time to understand what this means for you, your company, and your career.

1. The Hard Facts About Squidbleed

2. What is Squid – and why is this important?

Squid is one of the most widely used web proxies in the world. A proxy is – simplified – a mediator between your computer and the internet. When you surf in a company, school, university, or public WLAN, your traffic often goes through such a proxy. It can save bandwidth, block dangerous websites, or log usage.

Squid has been open source since the late 1990s, is used in millions of networks, and often runs on systems that are rarely updated – routers, embedded devices, legacy infrastructure. That's exactly what makes Squidbleed so dangerous: The bug has already been identified, but it will still exist in productive systems for years to come, because many Squid instances will never or only rarely be patched.

3. The 29-Year Time Capsule Effect: How could the bug have survived so long?

The story begins in January 1997. A developer wanted Squid to be compatible with old Novell NetWare FTP servers. These servers had the peculiarity of formatting directory lists with additional spaces. The developer therefore added code that skips over spaces when parsing FTP directory lists.

The technical cause of the error is a classic from C programming: The function strchr searches for a specific character in a character string. When called with the null terminator (the end of the string) it returns – according to the C standard – not NULL, but a pointer to the end. The loop in the Squid code therefore runs over the end of the actual string and reads memory areas that just happen to be there – often HTTP requests from other users.

The fix: just one additional condition in the loop. In pseudo-code:

// Before (vulnerable):
while (strchr(w_space, *copyFrom)) ++copyFrom;

// After (safe):
while (*copyFrom && strchr(w_space, *copyFrom)) ++copyFrom;

Just one extra *copyFrom &&, a simple null check. Three decades of code reviews, security audits, and academic analyses – no one saw it.

Why no one found the bug

There are three reasons why Squidbleed has been slipping through for 29 years:

1. The bug is in legacy code for legacy protocols. NetWare FTP is hardly used anywhere today. Who searches in an obscure code branch for errors?
2. The bug is harmless – until it's not. The code works correctly for 99.99% of inputs. Only in a very specific constellation does the leak occur.
3. Human attention is limited. Squid has around 200,000 lines of C code. Critically examining each line 30 years in a row is practically impossible – for humans.

4. How Claude Mythos found the bug

Rong, a researcher at Calif.io, told the story like this: He was sitting in a plane with a 10-year-old Squid proxy on its onboard WLAN. Out of curiosity, he fed Claude Mythos Preview with the Squid source code and asked the AI to examine the FTP parser for potential vulnerabilities. Within a few minutes, the model pointed out the suspicious strchr construction and explained precisely why it is flawed.

The reason Claude Mythos finds the bug where humans fail: The model has the complete C programming standard, all relevant CVE databases, and millions of code examples as references in context. It doesn't see the suspicious line as "this is how it grew," but immediately compares it to hundreds of similar constructs where this error was catastrophic.

A crucial addendum: Honor where honor is due

After the publication, it was found that another security researcher – Pavel Kohout from Aisle – had independently discovered and reported the bug on March 4, 2026, before Calif.ios Report in April. This does not diminish the significance of the AI find – but it shows that the error search in critical infrastructure is now running in parallel at multiple locations, with and without AI support.

5. The Patch Twist: Why Many Admins Think They Are Safe – and Aren't

In addition, Linux distributions often ship their own, older Squid versions. Debian, for example, still provides Squid 5.7. If you rely solely on the version number, you may overlook Squidbleed. Calif.io's recommendation: Check directly in the source code of the file FtpGateway.cc if the patch is included – or alternatively, completely disable the FTP protocol.

6. Why Squidbleed is the Bigger Story

Squidbleed may be a relatively moderate bug (CVSS 6.5) at first glance. The real reason the security community is paying attention: It's not just one find. It's a pattern.

Anthropic Project Glasswing – 23,000 Bugs in Months

Squidbleed is just one of the most prominent findings. Anthropics Project Glasswing has identified around 23,000 security vulnerabilities in open-source projects within a few months. OpenAI's competitor initiative is called Patch the Planet and works on similar scales.

The Consequence

The security research sector is undergoing a revolution. What was once laborious, specialized work is being transformed into assembly-line work by AI agents. Positive consequence: more bugs are found, systems are safer. Negative consequence: the same techniques can also be used by attackers.

The Five Eyes Warning

The Five Eyes intelligence agencies (USA, UK, Canada, Australia, New Zealand) refer to this in their June 2026 report: Do not expect years, but months until Frontier AI cyberattack capabilities are "fundamentally transformed". Malicious actors will work with the same tools – without responsible disclosure.

7. What this means for you professionally

Whether you work in IT or not, three consequences will affect your career:

For IT and security professionals

The half-life of pure "bug-finding" skills is decreasing. To remain valuable, you must develop towards context understanding, prioritization, communication, and risk assessment. The art is no longer "which bug can I find?", but "which of the 23,000 found bugs are really critical for this company?".

For software developers

Code Review will be supported by AI. In 2027, as a senior developer, you will be expected to use AI tools naturally and know their limitations. The skill "I write clean code that humans and AI can review together" replaces "I am very good at Python alone".

For everyone: Basic understanding will become more valuable

Squidbleed is a simple story on a meta-level: A loop forgot to check if it had reached the end of the character string. Those who understand this meta-level – without having to program themselves – will be able to make more targeted decisions in 2026 about which reports and warnings to trust. Basic IT understanding will become a general career competence.

8. Conclusion: Squidbleed is a wake-up call, not an isolated incident

Squidbleed will not be the last 29-year-old bug that AI models will expose in 2026. It is likely to be one of the first of thousands. For you as a professional, this means that the half-life of IT knowledge will continue to decrease. The half-life of meta-skills – understanding of systems, critical thinking, clear communication – will increase.

Exactly what Skill Tandem is counting on. On our platform, you'll find learning partners and mentors with whom you can build meta-competencies: from IT basic understanding to cybersecurity basics to strategic thinking. Learning in tandem helps you stay connected in a rapidly accelerating tech world – without drowning in a sea of detail information. Sign up for free now and start learning with a partner!

FAQ: Frequently Asked Questions about Squidbleed

Is my home WLAN at risk?

Almost certainly not. Squid Proxy is used in company, educational, and public networks, not in standard home routers. However, if you are connected to the company network using a company laptop, you may be indirectly affected.

Is Squidbleed as bad as Heartbleed?

No, much less severe. Heartbleed (2014) had a CVSS of 7.5 and directly affected encrypted HTTPS connections. Squidbleed is at 6.5 and only affects under specific configurations and not standard HTTPS connections. Nevertheless: In company setups with TLS interception, the impact is real.

Can Squidbleed already be exploited?

Yes. Proof-of-Concept exploit code has been publicly available since June 2026. Attackers with access to an internal network or a public WLAN with Squid Proxy can realistically exploit the bug. Therefore: patch quickly.

How quickly can AI models find something like this?

Claude Mythos needed only a few hours of targeted analysis for Squidbleed, according to Calif.io. However, the model was directed at the relevant code area from the start. For a blind scan of a complete codebase, current models still take significantly longer – trend: rapidly decreasing.

Does this mean that open-source software is less secure than commercial software?

No, on the contrary. Open Source is simply more scrutinizable. Commercial closed-source software likely has comparable (or more) legacy bugs, but they cannot be publicly investigated.

How can I professionally adapt to this trend?

Invest in combination skills: Understand how AI finds bugs, how to prioritize security reports, and how to prioritize risks. A learning partner on Skill Tandem helps you dive deeper into these topics with others – often faster and more sustainably than in solo online courses.