IIRC from the Anthropic report, the alleged danger of Mythos isn’t that it finds more vulnerabilities than previous models, but that it’s significantly more successful at exploiting them. Which this doesn’t seem to test.
IIRC from the Anthropic report, the alleged danger of Mythos isn’t that it finds more vulnerabilities than previous models, but that it’s significantly more successful at exploiting them. Which this doesn’t seem to test.
I would naively expect finding and exploiting to be related. Leaving this comment so someone can correct it, which would be interesting.