Context: earlier reporting on rising concern about using advanced computing for offensive research and patch management remains a critical defense focus.

Anthropic's Mythos Preview can rapidly weaponize newly disclosed software vulnerabilities, turning patches into working exploits within hours, according to Anthropic research shared with Axios.

Claude Mythos Preview can transform an attack that historically took weeks into a matter of hours, raising concerns about defense timelines and patch deployment speeds.

Mythos Preview can quickly generate working exploits for patched vulnerabilities, creating a patch gap where unpatched machines remain at risk.

Other open-source models and competing systems are finding bugs at similar levels, not just Mythos.

Testing on Windows used patched binaries, decompiler output, and public vulnerability data due to lack of source code, making it a tougher test than Firefox.

The findings come amid heightened government scrutiny of AI and cybersecurity, including a White House executive order to assess national security risks of AI and strengthen oversight.

Most cyberattacks target known vulnerabilities, and patching often requires testing and downtime before deployment, underscoring the ongoing patch management challenge.

Frontier red team tested Mythos against vulnerabilities in Mozilla Firefox and the Windows kernel disclosed in early year disclosures, including post-cutoff vulnerabilities.

Broader implications: open-source and commercial AI systems are advancing vulnerability research, raising concerns about dual-use capabilities in offensive security.

Costs observed: PoCs for Firefox were produced within a three-million-token budget, while full Windows privilege-escalation exploits cost about $2,000 each, signaling low weaponization barriers.

The practical implication is that the cost and time to weaponize patches have dramatically decreased, expanding the pool of potential N-day attackers to include non-experts with limited API access.