@AnthropicAI blog post: https://red.anthropic.com/2026/n-days/

@logangraham @AnthropicAI It's always 66% faster efficient than the previous in your testing.

"Mythos Preview is a step change improvement in turning a crash into a full exploit"

25 years and FSD still can't drive itself. Cool step change but you gave it the test answers and it got a C.

@gbrl_dick @matrosov @AnthropicAI we didn't, but I can try to follow up on that. (We focused on time because the thing people are concerned about is how much time an unpatched system is vulnerable to the n-day derived from the patch... the earlier it's found, the more risk)

@logangraham @matrosov @AnthropicAI this is pretty nuts. will check the paper but did you happen to publish a token x axis?

@logangraham @AnthropicAI This is good, mostly because I’m so bored of all the red team stuff

@logangraham @AnthropicAI more n-day blue team defense posts would be super helpful!

@logangraham @AnthropicAI The x axis being time and not tokens means that mythos is actually hella token efficient if it still finds stuff faster while likely having way slower tps

@logangraham @AnthropicAI Thanks for sharing your team's work publicly with the infosec community!

@logangraham @AnthropicAI 👏👏👏

@logangraham @AnthropicAI is there an internal code name for the ultimate mythos red-team orchestrator/agent yet? that line that continues trending upwards is vv interesting after 14+ hours.

@logangraham @AnthropicAI the chart shows a clear gap

mythos pulling ahead significantly on both speed and total exploits reproduced on real windows kernel cves is one of the more concrete agentic capability jumps we've seen

@axios More Anthropic hype please 🤮

@logangraham @AnthropicAI The most important shift may not be that AI is getting better at finding vulnerabilities. It's that the time between disclosure and exploitation is collapsing. In cybersecurity, reducing that window changes the economics of defense entirely.

@logangraham @AnthropicAI In a year Mythos will look trivial is probably the most important line in this post.

@logangraham @AnthropicAI PURPLE team is where it’s at 😎💬 let’s work together

@logangraham @AnthropicAI Most companies would bury this finding. Anthropic published it and said start preparing. That is a different kind of transparency

@logangraham @AnthropicAI Publishing this is important precisely because of how uncomfortable it makes people. The gap between 'AI can help write code' and 'AI can reproduce CVEs in hours for $1k' is collapsing faster than most security teams realize. Responsible disclosure > silence.

@logangraham @AnthropicAI looks like a figure from a paper introducing new RLVR method

@logangraham @AnthropicAI let the hype begin

@logangraham @matrosov @AnthropicAI oh no i get it — makes sense.