The Horizon3 disclosure of a vulnerability in Rejetto HFS was followed within days by exploitation attempts against vulnerable hosts, including activity sourced from a China-hosted IP. The sequence has drawn attention because the vulnerability was identified and demonstrated with help from Anthropic’s Mythos model, yet the available telemetry does not establish who operated the probing infrastructure or whether any victim was compromised.
The incident illustrates two separate developments that should not be collapsed into one claim. First, AI-assisted vulnerability research produced a disclosure. Second, public monitoring recorded rapid attempts to target exposed systems after the disclosure. The location of a source IP can describe network telemetry, but it cannot by itself identify an operator, motive, or state connection.
Mythos Helps Demonstrate the HFS Authentication Bypass
Horizon3’s Zach Hanley disclosed CVE-2026-61500 on September 30 and credited Anthropic’s Mythos model with helping identify and demonstrate the issue. Horizon3 described a session-forgery and authentication-bypass path rooted in predictable material derived from JavaScript Math.random(). The condition can lead to administrative access and remote code execution in affected HFS installations.
The technical work described by Horizon3 combined reversible V8 xorshift128+ output, a leaking application path, and the Z3 solver in the proof of concept. The research connected a weakness in session material to a route toward privileged access, while also showing how an AI system could assist a human researcher in demonstrating the risk.
The card for this incident does not authorize claims about Mythos independently discovering the vulnerability, operating without human direction, or creating a complete attack campaign. The verified account is narrower. Horizon3 credited the model with helping identify and demonstrate CVE-2026-61500, and Horizon3 supplied the technical description of the weakness. That distinction is important as security teams assess what AI-assisted research means for disclosure speed and defensive workload.
Rejetto’s v3.2.1 release is dated July 13 and says prior versions contained vulnerabilities that could allow administrative access. The release history provides a remediation reference, but the verified facts do not establish how many systems remained exposed after the fix, how widely the vulnerable versions were deployed, or whether a particular organization suffered a confirmed breach.
This episode also fits into a wider measurement problem in AI-enabled security research. The Register reported that Patrick Garrity’s tracker listed 286 CVEs uncovered by Mythos and Project Glasswing as of October 2. That figure is a tracker count reported by The Register, not a measure of successful exploitation, financial damage, or confirmed victim impact. It also does not show that every listed CVE followed the same workflow as CVE-2026-61500.
China-Hosted Telemetry Arrived After Disclosure
The timing of the observed activity is the central operational fact. The Register reported exploitation attempts by October 1, shortly after the September 30 disclosure. VulnCheck researcher Patrick Garrity said a China-hosted IP targeted vulnerable hosts in the United States and Japan. The report also described later hits from U.S. IPs that appeared to be proxy-related.
Those observations support a statement about source-IP telemetry, not an attribution finding. The verified facts do not identify the operator as Chinese, China-linked, or state-backed. IP telemetry is not confirmed actor identity.
The Register’s account does not establish successful compromise, persistence, victim impact, or a breach. It also does not establish actor identity. Maintaining those boundaries preserves the difference between telemetry and attribution.
For defenders, the practical response is more concrete than the attribution question. Organizations running HFS should identify versions in use, review exposure, apply the available release guidance, and examine logs for suspicious authentication or administrative activity. The supplied facts do not provide a universal remediation timetable, an affected-installation count, or a confirmed list of victims, so security teams should avoid treating the public telemetry as a complete incident inventory.
AI-Assisted Disclosure Raises Attribution and Defense Questions
The incident combines a fast-moving disclosure cycle with a familiar challenge in cyber investigations: infrastructure evidence is not identity evidence. Horizon3’s technical account describes how predictable randomness, application-path leakage, and solver-assisted analysis contributed to the proof of concept. The Register’s reporting adds the timeline of exploitation attempts and the China-hosted source-IP observation. Each source answers a different question, and neither source resolves who was behind the traffic.
That separation also matters for the US-China technology contest. A China-hosted IP touching vulnerable U.S. and Japanese hosts may attract immediate strategic interpretation, but the verified facts support only the narrower network observation. Treating hosting geography as attribution could misstate the evidence and make future incident reporting less reliable. The responsible conclusion is that exploitation telemetry crossed borders quickly after disclosure, while the operator remains unresolved.
The verified record supports an AI-assisted vulnerability disclosure, rapid attempted exploitation, a China-hosted source IP targeting vulnerable U.S. and Japanese hosts, and later U.S. hits that appeared proxy-related. It does not support a confirmed actor identity, a state-backed allegation, a confirmed compromise, or a measured victim impact.
For this HFS case, attribution has not been solved. The record supports an AI-assisted disclosure and rapid probing, while stronger evidence would be required before assigning responsibility.
For related EastFrontier context, see Anthropic Tests Z.ai’s GLM-5.3 on Advanced Cyber Tasks.