
Horizon3.ai researchers used Anthropic’s Mythos model to find a cryptographic weakness in Rejetto HTTP File Server (HFS), tracked as CVE-2026-61500. The flaw lets an unauthenticated attacker forge admin session cookies and then run code on the server.
Anthropic launched Project Glasswing to secure the world’s most critical software. Horizon3 joined in July 2026 and has used Mythos in its vulnerability research pipelines. Rejetto HFS is an open-source file-sharing tool.
Anthropic Mythos AI Finds Rejetto HFS Flaw
It was previously listed in CISA’s Known Exploited Vulnerabilities catalog for CVE-2024-23692, a template injection that led to remote code execution. HFS 3.x was rewritten in TypeScript, and Horizon3 analyzed that version with a custom harness.
According to research published by Horizon3.ai, HFS builds its session-signing key from Math.random(). The Koa web framework then uses keygrip to sign all session cookies with that key. Node.js’s V8 engine implements Math.random() with the xorshift128+ algorithm.
That algorithm is non-cryptographic, and its outputs can be reversed. An attacker who observes enough outputs can recover the internal state and work backward to earlier values.
Mythos found a second flaw that supplies those observations. During login, the loginSrp1 function stores a raw Math.random() value as a session ID. The koa-session cookie is base64-encoded JSON that is signed but not encrypted.
The client can therefore read the exact 52-bit double. Reaching this step requires only a valid, login-enabled username, not a password.
A verifying agent confirmed the finding. It noted that the code comment “randomness at start gives some extra security” shows the key was meant to be secure.
Mythos also noticed that HFS’s administrative API supports custom endpoints that can execute arbitrary JavaScript. That turns the authentication bypass into code execution.
The researchers describe the steps as follows: confirm the built-in admin account exists using a separate user-enumeration flaw, which Mythos also found and which was disclosed; sample the unauthenticated endpoint that leaks raw Math.random() values 12 times; feed the values into the Z3 SMT solver as constraints to recover the xorshift128+ state; step the state backward to rebuild the signing key generated at startup; forge a valid admin session cookie; confirm the cookie grants authentication; and use built-in HFS functionality to run commands on the server.
The forged session includes the allow_session_ip_change flag, which bypasses IP binding. Default settings also impose no admin network restriction. The attacker’s own legitimately issued cookie and HMAC signature serve as an offline oracle to verify each candidate key.
Horizon3 said Mythos recognized that the weak PRNG and the leak formed a chain. It then wrote a working proof of concept with the Z3 solver and demonstrated arbitrary command execution, all without follow-up prompting.
The researchers said they had abandoned similar cryptographic bugs before because they lacked the math expertise and the work was not economical. They also said they had never seen an SMT solver used against a real application this way.
Horizon3 warns that cheaper automation will change which bug classes attackers weaponize at scale. Complex, less reliable flaws such as memory-safety bugs may become practical targets. The firm added that off-the-shelf harnesses with current models can surface findings in projects that have never had a security review.
Horizon3 did not list a patched version or a fix timeline in the material provided. Until you confirm a fix, take these steps: check the Rejetto HFS GitHub repository and advisories for a patched release, restrict access to HFS admin interfaces by network, avoid exposing HFS directly to the internet, monitor for repeated loginSrp1 requests and unexpected admin sessions, and review custom endpoints and server logs for unauthorized changes.
Horizon3 plans to detail further Mythos-discovered vulnerabilities as they are patched.
The post Anthropic Mythos AI Finds Rejetto HFS Flaw That Lets Attackers Forge Admin Sessions and Execute Code appeared first on Cyber Security News.
Discover more from RSS Feeds Cloud
Subscribe to get the latest posts sent to your email.
