Why Goldman Sachs is 'hyper-aware' of Anthropic Mythos AI risks

Mythos AI autonomously found a 27-year-old vulnerability in OpenBSD that had been overlooked in millions of automated tests

Goldman Sachs CEO David Solomon has revealed that the bank is in close collaboration with Anthropic following startling warnings regarding the cybersecurity threats posed by its latest AI models, Claude Mythos.

As a part of a broader strategy to defend against evolving hacking tactics, the US bank has constantly been involved in intense evaluation of large language models (LLMs).

"Obviously the LLMs are making rapid progress and we're hyper-aware of the enhanced capabilities of these new models with the help of the US government and the model publishers," Solomon told analysts on an earnings call on Monday.

According to Solomon, Claude Mythos AI possesses cutting-edge capabilities, known for identifying critical cyber vulnerabilities in various operation systems and web browsers.

"We're working closely with Anthropic and all of our security vendors to kind of harness frontier capabilities wherever it's possible. And this will continue to be an important focus," he added.

Solomon further assured, "We are very focused on supplementing our cyber and infrastructure resilience. And this is part of our ongoing capabilities that we have been investing in, and are accelerating our investment in."

Earlier this week, the US Treasury Secretary Scott Bessent and Federal Reserve Chair Jerom Powell issued a dire warning to the financial sector, especially bank CEOs regarding AI-driven cyber risks.

In the meetings, the heads of banks were instructed to defend their financial systems against rising AI-powered cyber threats.

Another warning also came from the UK government's AI Security Institute (AISI) on Monday, stating Mythos "was a step up over previous AI models in terms of the cyber threats it posed."

According to AISI, Mythos is highly capable of performing multiple tasks and discovering loopholes in IT systems without human oversight.

Surpassing human professionals, Mythos made history as the first AI model to conquer a 32-step cyber-attack simulation designed by the AISI, achieving a success rate of 30% across 10 attempts.

AISI also reported that during the testing Mythos appears to be capable of autonomously targeting small and weak IT systems. The institute also issued a dire warning that in near future AI models will only improve, surpassing Mythos. Therefore, investment in cyber defence has become the need of the hour.

Given Claude Mythos unprecedented cyber capabilities, Anthropic has decided not to release this model publicly. Under its newly-launched "project glasswing" project, Mythos will be available to Google, OpenAI, and Microsoft and other organizations.