Anthropic said it will withhold public release of its newest AI model, Mythos Preview, citing “unprecedented” abilities to find and chain software vulnerabilities—capabilities the company fears could be weaponized at scale. The move, which limits access to select partners while they harden defenses, underscores rising concern that generative AI could shift the cyber balance of power toward attackers. The Treasury Department convened major financial institutions to discuss the accelerating risks, according to an agency spokesperson.
Security researchers warn that advanced coding models can rapidly identify flaws and assemble complex exploits, lowering the skill bar for ransomware groups and state-linked actors. Anthropic’s Logan Graham said similarly capable models could be broadly available within 6–12 months, including from overseas competitors. Experts flagged elevated risk to critical infrastructure and sectors with low tolerance for downtime, such as health care and manufacturing, though widespread, catastrophic failures remain unlikely due to segmented and offline systems.
The episode highlights a maturing AI safety debate: how to manage cutting-edge capabilities that can boost both defense and offense. With U.S. officials tracking Iranian probes of utilities and energy firms, companies face an intensifying arms race—one where defenders must be right every time and attackers only once.
AI ‘Vulnpocalypse’: How powerful models could hand hackers the upper hand
- Trending
- Comments
- Latest
