Advanced artificial intelligence systems are rapidly changing the cybersecurity landscape, with new research showing that powerful AI models can transform publicly disclosed software vulnerabilities into working cyber exploits in a matter of hours rather than weeks.
Anthropic's newly unveiled Mythos Preview model successfully generated working exploits for software flaws in Microsoft Windows and Mozilla Firefox during internal testing, raising concerns that cybercriminals could soon weaponize known vulnerabilities at unprecedented speed, Axios reported Monday.
The findings suggest the biggest AI cyberthreat may not be discovering new bugs but dramatically accelerating the exploitation of flaws that organizations have not yet patched.
In tests conducted by Anthropic's frontier red team, Mythos generated its first proof-of-concept exploit for a Windows kernel vulnerability in just 31 minutes.
Across 21 Windows kernel vulnerabilities, the model triggered system crashes known as the "blue screen of death" in 18 cases and produced eight distinct exploits.
The most complex exploit required about 5.7 hours to develop.
The model also demonstrated success against Mozilla Firefox.
According to Axios, Mythos created eight working code-execution exploits across 18 security patches.
Cybersecurity experts have long warned that most successful cyberattacks target known vulnerabilities rather than previously undiscovered flaws.
Organizations often require days or weeks to test and deploy software updates, creating a window of opportunity for attackers.
Anthropic estimated that Mythos generated Windows privilege-escalation exploits at a cost of roughly $15,700 in computing resources, or about $2,000 per exploit.
The concerns extend beyond Anthropic. The company told Axios that some open-source AI models are already approaching the cyber capabilities of leading frontier systems, potentially putting advanced offensive tools into the hands of a wider range of actors.
The development comes as Anthropic expands cooperation with the U.S. government.
The Financial Times reported Thursday that the company is helping the National Security Agency deploy Mythos for cybersecurity and national security applications, with engineers reportedly embedded within the agency to customize the technology for specialized missions.
According to the report, supporters of the effort argue that the United States must develop advanced AI-powered cyber capabilities because adversaries such as China and Iran are likely pursuing similar technologies.
The growing capabilities of AI systems have prompted action from the Trump administration.
President Donald Trump recently signed an executive order establishing a voluntary framework for AI security reviews and directing federal agencies to develop methods for evaluating the cyber capabilities of advanced AI models.
Supporters say such measures will help ensure the United States remains ahead of foreign adversaries while addressing the national security risks posed by increasingly powerful artificial intelligence systems.