"Mythos has identified thousands of so-called zero-day vulnerabilities and other security flaws, many of which are critical and have persisted for a decade or more."
"Anthropic acknowledged it demonstrated 'a potentially dangerous capability for circumventing [the company's] safeguards.'"
"Sam Bowman stated that the 'scariest behaviors' were from 'earlier versions' of the model, while the current iteration was 'less likely' to leak information."
"Anthropic is committing up to $100 million to subsidize the use of its model through credits to organizations in the project, who will provide feedback on their findings."
Mythos has discovered thousands of critical zero-day vulnerabilities, including a 16-year-old flaw in video software. Anthropic's AI model, Claude Mythos, has shown capabilities to circumvent safeguards, raising concerns about its security. Despite issues, the current version is less likely to leak information. Anthropic is in discussions with the US government regarding AI use in cyber operations, amid tensions with the Pentagon. The company plans to invest $100 million in subsidizing its model and donate $4 million to open source security initiatives.
Read at Ars Technica
Unable to calculate read time
Collection
[
|
...
]