Anthropic said on Tuesday it will not publicly release Mythos, a frontier-AI model capable of automatically discovering software vulnerabilities [1, 2].
The decision highlights a growing tension in the AI industry between developing powerful defensive security tools and preventing the creation of weapons for cyberattacks. Because Mythos can identify flaws at scale, the company said the tool could be weaponized by malicious actors if released openly [1, 4].
Mythos is currently available only to a limited set of partners for testing [2, 7]. The model has already demonstrated a high level of precision in identifying deep-seated errors. In one instance, Mythos flagged a 27-year-old software vulnerability that had survived decades of manual code review [3].
Experts are divided on whether the model represents a net gain for digital security. Some view the system as a critical defensive asset that can patch holes before hackers find them [5]. Others said the model is so powerful at revealing software vulnerabilities that the risk of it falling into the hands of bad actors outweighs the benefits of public access [4].
Anthropic, the developer of the Claude assistant, is positioning Mythos as a frontier-AI model [1]. This classification refers to highly advanced systems that can exhibit general-purpose capabilities across many domains. The company's refusal to release the model publicly follows an announcement this week regarding the system's capabilities [1, 2].
While the model remains restricted, its existence shifts the debate over AI's role in cybersecurity. The ability to automate the discovery of vulnerabilities could potentially accelerate the pace of software patching, or accelerate the speed at which attackers can find entry points into secure systems [3, 5].
“Mythos flagged a 27-year-old software vulnerability that had survived decades of manual code review”
The restriction of Mythos signals a shift toward 'closed-door' development for AI models with high dual-use potential. By limiting access to a small group of partners, Anthropic is attempting to reap the benefits of vulnerability discovery without providing a roadmap for hackers. This move may prompt other AI labs to adopt similar safety guardrails as models move from generating text to actively interacting with and auditing complex codebases.
