CVE-2026-61500: Anthropic Mythos-Found Flaw in Rejetto HFS Under Active Attack Within 24 Hours

A critical authentication-bypass flaw in Rejetto HTTP File Server came under active exploitation on October 4 — less than 24 hours after Wednesday's public disclosure. If you run HFS, update to version 3.2.1 immediately. The flaw was found not by a human researcher working alone, but by security researcher Zach Hanley using Anthropic's Mythos AI model, marking one of the first confirmed cases of an AI-assisted CVE going from discovery to mass exploitation in under a day.

The Flaw

Rejetto HFS (HTTP File Server) is a lightweight, free file-sharing tool widely used on Windows by small businesses, home labs, and individual users who want a simple way to serve files over a local network or the internet. CVE-2026-61500 is an authentication bypass — meaning an attacker can gain access to the server without a valid username or password.

The mechanism is subtle but reproducible. HFS derives its session-cookie encryption key from JavaScript's Math.random() function. That function is not cryptographically secure — its output is deterministic and reversible if you know the seed. A separate code path in HFS leaks enough Math.random() output to reconstruct the seed. Combined, these two flaws allow an attacker to forge a valid session cookie and authenticate as any user, including the administrator.

Hanley used Anthropic's Mythos — a model Anthropic has positioned as specialized for mathematical reasoning and vulnerability research — to identify the connection between the two code paths. According to The Register's reporting, Mythos flagged the Math.random() seed-recovery angle as the critical link that turned a theoretical information-leak into a practical authentication bypass.

Active Exploitation in Under 24 Hours

The CVE was disclosed publicly on October 3. By October 4, VulnCheck's canary infrastructure — honeypot servers instrumented to detect attacks — had logged active exploitation attempts. The attacking IP addresses were traced to China-based infrastructure, with U.S. and Japanese HFS deployments as the primary targets.

The speed of exploitation is notable but not unprecedented for high-value, low-complexity flaws. An authentication bypass that requires no prior credentials and runs against a widely deployed tool is exactly the kind of vulnerability that automated exploit frameworks pick up immediately after disclosure.

What This Means for Defenders

Short term: Patch now. Rejetto has released HFS 3.2.1, which replaces the Math.random()-based key derivation with a cryptographically secure pseudorandom number generator (CSPRNG). The update is available on the Rejetto GitHub repository.

If you cannot patch immediately, the interim mitigations are:

• Place HFS behind a VPN or IP allowlist so it is not exposed directly to the internet
• Rotate all existing session tokens (restart the HFS service)
• Monitor logs for unexpected authentication events

Longer term: CVE-2026-61500 is an early data point in a larger question: what happens to the vulnerability disclosure ecosystem when AI models can find flaws faster than human researchers? The traditional 90-day responsible disclosure window assumes a race between a researcher notifying a vendor and a separate attacker independently discovering the same flaw. If AI models compress discovery timelines on both sides, that 90-day window may need to shrink.

The Mythos Angle

Anthropic's Mythos model has not been publicly released. It has been described by Anthropic researchers as a specialized system optimized for mathematical reasoning and structured vulnerability analysis — distinct from Claude's general-purpose design. Hanley's use of Mythos to find CVE-2026-61500 is one of the first public demonstrations of the model's capabilities in a real-world offensive security context.

The finding raises a straightforward question that the security industry has not yet answered cleanly: when an AI model finds a vulnerability, who bears disclosure responsibility — the researcher who ran the query, the lab that built the model, or both?

What to Watch

Whether Anthropic clarifies its policy on Mythos-assisted vulnerability disclosure, and whether CISA (the U.S. Cybersecurity and Infrastructure Security Agency) adds CVE-2026-61500 to its Known Exploited Vulnerabilities catalog — which would trigger mandatory patching timelines for federal agencies. Given the confirmed active exploitation, that listing is likely.