AI-found Rejetto HFS flaw exploited within a day of write-up

The gap between disclosure and attack has shrunk to roughly a day, and AI-found bugs are now part of that race. Attackers began exploiting CVE-2026-61500, a critical flaw in Rejetto HTTP File Server (HFS) discovered with Anthropic's Mythos model, about 24 hours after a public technical write-up appeared.

The vulnerability carries a CVSS score of 9.3. According to SecurityWeek, HFS used JavaScript's Math.random(), which runs on the reversible xorshift128+ generator, to create session-cookie values. An attacker who collects login responses can reconstruct the generator's state, recover the signing key, forge administrator session cookies and gain remote code execution.

Horizon3.ai researcher Zach Hanley found the flaw in June 2026 using Mythos. Horizon3.ai said the model used "advanced mathematical reasoning to recognize that Math.random() PRNG outputs could be reversed to reconstruct the secret session-cookie signing key." Rejetto fixed the problem in HFS 3.2.1 on 13 July. Its advisory said: "Multiple security vulnerabilities have been found in all previous versions, potentially allowing an attacker to gain administrative access to HFS."

Horizon3 then published its technical write-up. VulnCheck saw exploitation attempts on 2 October, within about 24 hours, from a China Telecom IP address against canary systems in Japan and the US, according to DEV Community. Four US-based IP addresses followed. The only fix is to upgrade to HFS 3.2.1 or later.

The Register called this the second Anthropic-linked vulnerability exploited in the wild, under a headline saying Mythos is "hardcore good at math". Mythos is Anthropic's restricted cyber-capable model. Google's Gemini 4 Argon launched to cyber defenders the same week.

The exposure falls on anyone still running an HFS version older than 3.2.1. The patch had been available since 13 July, so the attackers' window opened when the write-up landed, not when the fix shipped. Operators who waited on the patch are the ones now on the wrong side of a one-day clock.

For AI labs, the story cuts both ways. Mythos found a bug that required recognising a mathematical weakness in a random number generator, which is the kind of finding that sells restricted cyber models to defenders. Yet the same output, once published, was exploited almost at once. Watch how vendors and researchers handle the timing between patch, write-up and exploitation, and whether Google's defender-focused launch the same week turns that timing into a selling point.