Anthropic’s experimental cybersecurity model Mythos is demonstrating how quickly AI can accelerate vulnerability research — and how narrow the window between discovery and real-world exploitation is becoming.
Security researchers using Mythos discovered a critical authentication bypass in Rejetto HTTP File Server (HFS) that can lead to administrator access and remote code execution. The vulnerability, tracked as CVE-2026-61500, was publicly disclosed on Wednesday. By the following day, security company VulnCheck was already detecting exploitation attempts against vulnerable servers.
The episode provides an unusually clear example of both sides of AI-assisted security research: models are becoming capable of finding subtle vulnerabilities, while attackers remain ready to weaponize newly published findings almost immediately.
Mythos Connected Two Problems Humans Could Easily Miss
Horizon3.ai researcher Zach Hanley discovered the vulnerability while using Mythos through Project Glasswing, Anthropic’s program that gives selected security researchers access to its specialized bug-hunting model.
The weakness originates in the way HFS creates authentication sessions. The software generated a random value using JavaScript’sMath.random()and passed it to the Koa web framework, where it was ultimately used to sign session cookies.
UsingMath.random()for security-sensitive operations is already questionable because it is not designed as a cryptographically secure random number generator. But Mythos went further than simply flagging the weak randomness.
The model discovered that HFS also leaked outputs fromMath.random()through another part of the application. It then recognized that those leaked values could potentially reveal the internal state of V8’s xorshift128+ pseudorandom number generator.
That connection transformed a coding weakness into an authentication bypass.
Mathematics Became the Exploit
Mythos proposed using Z3, Microsoft’s open-source satisfiability modulo theories solver, to reconstruct the generator’s internal state from the exposed values.
Once the state could be recovered, an attacker could predict the random value used as the session-signing key. That would allow them to forge legitimate session cookies, impersonate an administrator and eventually execute code remotely.
The Horizon3 researchers said they could not recall seeing an SMT solver used in this particular way to exploit a cryptographic weakness in a real application.
What makes the finding notable is not simply that an AI spotted an insecure random-number generator. Mythos identified two weaknesses in different parts of the software, understood their mathematical relationship and constructed an exploitation strategy connecting them.
That is much closer to vulnerability research than conventional automated code scanning.
Exploitation Followed Almost Immediately
The speed at which attackers reacted is equally important.
VulnCheck detected exploitation attempts the day after technical details became public. Initial activity came from an IP address hosted in China and targeted vulnerable systems in the United States and Japan. Later attempts originated from two US addresses that appeared to be acting as proxies.
The location of those systems does not establish who was behind the attacks, particularly because threat actors routinely route traffic through compromised or rented infrastructure.
Rejetto has fixed the vulnerability, and users should upgrade HFS to version 3.2.1 or later.
AI Is Compressing the Vulnerability Timeline
Mythos remains restricted rather than publicly available. Anthropic has said the model is sufficiently capable that it should currently be limited to selected security researchers through Project Glasswing.
The results help explain that caution. According to VulnCheck researcher Patrick Garrity’s tracking, Mythos and Project Glasswing had been credited with discovering 286 CVEs by October 2.
Most had not been observed under active exploitation, but CVE-2026-61500 became the second known Mythos-linked vulnerability to attract real-world attacks.
The larger issue is therefore not whether AI can find vulnerabilities. Increasingly, it can.
The more consequential change is speed. AI can analyze large codebases, connect weaknesses across different components and apply mathematical techniques that might previously have required substantial specialist effort. Once the vulnerability becomes public, attackers can use the same generation of tools to understand and weaponize it.
Cybersecurity has always involved a race between discovery, patching and exploitation. AI is beginning to make every stage of that race faster — and CVE-2026-61500 shows how quickly defenders may now need to move.