Is Anthropic's New AI Model Too Dangerous? Mythos Could Let Hackers Attack Faster Than Humans

Anthropic, the California-based AI company that made Claude, is preparing to release its latest artificial intelligence model, Mythos, to select global tech and cybersecurity firms. But it's raising fresh concerns about the risks posed by AI in the wrong hands.

The model, designed to identify software vulnerabilities at unprecedented speed, will initially be available to firms including Amazon, Apple, Cisco, Google, JPMorgan Chase, and Microsoft, according to Anthropic.

The announcement came during Anthropic's Builder Summit in Bengaluru, India, earlier this month, when co-founder and chief executive Dario Amodei outlined the company's approach to balancing AI innovation with potential security threats. Mythos, officially branded Claude Mythos Preview, is not intended for public release due to its capacity to accelerate cyberattacks beyond human limits, Anthropic said.

The Risk Anthropic's 'Mythos AI' Enters

AI-driven cybersecurity has been a growing frontier in recent years, with both attackers and defenders racing to harness machine learning to probe and patch systems. Anthropic's Mythos represents a significant escalation.

Experts told CNN that AI agents of this scale could automate vulnerability discovery and exploitation far faster than hundreds of human hackers, representing what one researcher described as a 'sea change' in the field.

Logan Graham, who leads Anthropic's AI model defence team, told CNN: 'We did not feel comfortable releasing this generally. We think that there's a long way to go to have the appropriate safeguards.' He added that the firm is working closely with the US government, briefing senior officials across multiple agencies and offering support for official testing of Mythos' offensive and defensive capabilities.

Anthropic hopes that by sharing the model with companies responsible for billions of users, it can reduce the risk of critical vulnerabilities appearing in popular internet browsers and operating systems.

Other recipients of the preview include chipmakers Broadcom and Nvidia, the Linux Foundation, and cybersecurity firms CrowdStrike and Palo Alto Networks.

The Potential and the Peril

The model's capabilities have already attracted attention following a leaked blog post last month, which claimed Mythos is 'far ahead' of other AI models in cyber tasks.

According to the post, Mythos 'presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.'

Anthropic asserts that Mythos has already identified 'thousands' of previously unknown software vulnerabilities in recent weeks, though CNN has not independently verified that figure. Such vulnerabilities are notoriously difficult for human researchers to detect and are prized by both state-backed intelligence agencies and cybercriminals.

Cybersecurity experts caution that the gap between AI-assisted attackers and defenders could widen if organisations do not adopt similar capabilities. Gadi Evron, founder of AI security firm Knostic, explained: 'Unlike attackers, defenders don't yet have AI capabilities accelerating them to the same degree. However, the attack capabilities are available to attackers and defenders both, and defenders must use them if they're to keep up.'

Dangerous in the Hands of Dangerous People

While some concerns around Mythos' potential misuse have been described as exaggerated, the leak highlighted an uncomfortable truth. AI of this power could give attackers a significant advantage unless organisations invest in equally advanced defensive tools.

The stakes are particularly high because critical infrastructure, financial systems and vast amounts of personal data could be exposed to rapid, automated attacks. Systems that defend these assets have traditionally relied on slower human analysis, and AI acceleration changes that dynamic, experts say.

Logan Graham emphasised the urgency for preparation. 'If models are going to be this good — and probably much better than this — at all cybersecurity tasks, we need to prepare pretty fast. The world is very different now if these model capabilities are going to be in our lives,' he said.

Mythos is not being released to the public at this stage, in the hope that they can use it to strengthen software before attackers exploit similar tools. Instead, Anthropic launched a controlled early access programme called Project Glasswing, making the model available from early April 2026 to a select group of major technology and cybersecurity firms.

Anthropic has emphasised that this gated preview is designed to help those organisations 'strengthen defensive cybersecurity work' and share findings with the industry, while it continues developing safeguards before any release might be considered.