Anthropic's super bug-finding Mythos model is going into Cisco and IBM products

Anthropic has begun expanding the accessibility of its bug-hunting Mythos model by allowing Cisco and IBM to incorporate it into their security products. At a Melbourne, Australia event, Cisco unveiled that its AI Defense suite, used for identifying weaknesses in AI models, agents, and applications, will now incorporate Mythos for testing clients' setups.

Jeff Schultz, Cisco's senior vice president for portfolio strategy and marketing, stated that AI Defense will use Mythos to enhance its capabilities in detecting potential threats. Anthropic's head of applied AI for the Asia-Pacific, Natalie Mead, indicated that the company has a long waiting list of customers eager to utilize Mythos, and collaborating with Cisco will facilitate the fulfillment of this demand.

Initially made available through Project Glasswing, a program for vetted organizations testing their codebases, Anthropic has since broadened access via other controlled programs, a decision that makes sense given the model's potential misuse if in the wrong hands. The Cisco deal exposes Mythos' functionality to end-users without permitting them to utilize the model itself.

When asked about similar partnerships with other third parties, Mead declined to disclose such information. Shortly after Cisco's announcement, IBM declared it would also integrate Anthropic's model into its Concert tool, which utilizes AI to identify risks and vulnerabilities in applications. IBM aims to leverage Mythos to detect flaws that conventional scanners might overlook, evaluate their significance, and generate developer-approved fixes.

The company has also released version 5.2 of Mythos and Fable, claiming they generate outputs up to 30% faster while costing less at $8 per million input tokens and $40 per million output tokens. Users have the option to run both models in "fast mode," which increases token generation speed by up to 2.5 times but doubles the cost to $16 per million input tokens and $80 per million output tokens.

Anthropic asserts that extensive pre-release safety testing was conducted on Fable 5.2 and Mythos 5.2, including novel evaluations to assess specific concerning behaviors like sandbox escapes. They concluded that high-stakes risks from misalignment remain low and that the model performs as well or better than recent models on most alignment measures.