AI-discovered flaw in Rejetto HFS triggers active exploitation from China
Anthropic's Mythos AI model identified a critical authentication bypass in Rejetto HTTP File Server (HFS), now under active exploitation. The flaw, CVE-2026-61500, allows attackers to forge admin sessions and execute remote code by exploiting a weak random number generator. Vulnerability researchers confirmed that Mythos used advanced mathematical reasoning to reverse-engineer the session key, marking a significant shift in AI-driven security discovery. Exploitation attempts have been detected from Chinese IP addresses targeting US and Japanese hosts, prompting urgent patching to version 3.2.1.
Key points
- CVE-2026-61500, rated 9.3 CVSS, exploits a non-cryptographic Math.random() generator in Rejetto HFS versions 3.0.0 through 3.2.0.
- Anthropic's Mythos model identified the flaw by recognizing that V8's xorshift128+ algorithm outputs were reversible and leaked during login handshakes.
- VulnCheck detected exploitation attempts on October 1, 2026, originating from a China Telecom IP address targeting canary deployments in the US and Japan.
- A patch was released in July 2026 (version 3.2.1), but a proof-of-concept exploit was publicly released in late September 2026.
- Mythos and Anthropic's Project Glasswing have collectively uncovered 286 CVEs, with this being the first to face real-world exploitation.
Background
Anthropic launched Mythos 5.1 in September 2026 as a specialized model for cybersecurity and life sciences, offering lower costs than its predecessors. This incident is the first instance of a Mythos-discovered vulnerability being actively exploited in the wild, following a period where the model was primarily used for internal testing and trusted access programs.
How outlets are covering it
The Register emphasizes Mythos's mathematical prowess in solving the cryptographic flaw, while The Hacker News and SecurityWeek focus on the technical mechanics of the session forgery and the timeline of the patch versus the exploit. ET Datacenters highlights the broader trend of AI-driven vulnerability discovery, noting that Mythos and Project Glasswing have found 286 CVEs, with this being the first exploited one. All sources agree on the origin of the exploitation attempts from China and the urgency of updating to version 3.2.1.
Why it matters
This incident demonstrates the growing capability of AI models to identify and exploit complex cryptographic vulnerabilities, shifting the threat landscape from human-led to AI-assisted attacks. The rapid exploitation of a patched vulnerability highlights the need for organizations to prioritize patch management and monitor for AI-driven threat actors.
What to watch
Security teams should immediately update Rejetto HFS to version 3.2.1 or later to mitigate CVE-2026-61500. Organizations should also monitor for unusual login activity and consider implementing additional authentication controls. The incident may lead to increased scrutiny of AI models in cybersecurity and a push for more robust random number generators in software development.
- Anthropic's super bug-hunting model Mythos is hardcore good at math, as latest vuln under attack shows theregister.com
- Attackers Target Rejetto HFS Flaw That Enables Admin Session Forgery and RCE The Hacker News
- New Anthropic Vulnerability Exploited by a Cyber Actor in China for Attacks on the US and Japan nacionale.com
- Exploitation Hits Rejetto HFS Vulnerability Discovered by AI SecurityWeek
- Anthropic's Mythos AI uncovers critical vulnerability in HTTP file server, exploited in the wild ET Datacenters
Want the full story? Read the original reporting
Read on theregister.com