AI-Assisted Research Helps Fix a 29-Year-Old Squid Proxy Flaw Before Attackers Found It
Researchers used AI to help uncover and responsibly disclose Squidbleed, a memory-leak flaw lurking in the Squid web proxy since 1997 — and the coordinated fix shipped before any known exploitation.
A Decades-Old Bug, Caught the Right Way
Every so often a security story comes along that's actually reassuring, and this is one of them. On June 22, 2026, researchers detailed Squidbleed (tracked as CVE-2026-47729), a memory-leak flaw in the widely used Squid web proxy that had been quietly present since 1997 — nearly 29 years. The genuinely good news is buried in how it was handled: it was found with the help of AI, responsibly disclosed, and patched before any known exploitation. That's the system working as designed.
Let me walk through what happened, because the process here is a small masterclass in defensive security.
What Squidbleed Actually Was
In simple terms, Squidbleed was a Heartbleed-style flaw — the kind where a program can be coaxed into reading slightly *past* the chunk of memory it's supposed to, and handing back whatever happened to be sitting there. The bug lived in Squid's FTP parser, and in the wrong conditions it could expose another user's uncleared HTTP request data: think credentials, session tokens, or API keys that lingered in memory.
Where the Risk Was Concentrated
Context matters for assessing real-world risk, so here's the honest scope. The flaw was most relevant in shared-proxy environments — corporate networks, schools, and public Wi-Fi setups where many users' traffic flows through the same Squid instance. Importantly, HTTPS traffic handled through opaque CONNECT tunnels was not affected, which meaningfully limited the blast radius. And for anyone needing an interim mitigation, disabling FTP support closed the door.
Why the AI Angle Is Encouraging
Here's the part I find most exciting from a defender's perspective. The flaw was uncovered with the assistance of Anthropic's Claude (Mythos) model, used to help review code that humans had been walking past for nearly three decades. This is the constructive face of AI in security: not flashy, just a tireless second set of eyes that can comb through old, gnarly code and surface latent issues so they can be fixed.
That's a real shift. Some of the most dangerous vulnerabilities are the quiet, ancient ones hiding in software everyone trusts. AI-assisted review gives the defensive side a powerful new tool for finding those needles before anyone with bad intent does.
Responsible Disclosure Did Its Job
The disclosure timeline is the other hero of this story. Rather than being dumped into the open, the issue was reported through coordinated disclosure, the patch was merged into Squid v8 in April 2026, and the fix shipped in v7.6 in June 2026. By the time the public learned about Squidbleed, the remedy was already available. That ordering — fix first, then talk — is exactly what protects users.
The Takeaway for Defenders
If you run Squid anywhere, the practical step is simple: update to a patched release (v7.6 or later), and apply the FTP mitigation in the meantime if you can't patch immediately. But the broader lesson is the uplifting one. A decades-old flaw was found by pairing skilled researchers with AI tooling, handled responsibly, and closed before causing harm. That's defense getting *faster*, and it's precisely the kind of story that makes me optimistic about where AI security is heading.
Sources: SecurityWeek — "Decades-Old Squid Proxy Flaw 'Squidbleed' Can Expose User Data" — June 22, 2026; Squid Project — security advisory and release notes (v7.6) — June 2026.