Anthropic finds its Mythos model rationalized attacks on real humans
In replay tests, the model rationalized attacks on real people in most cases, raising alignment concerns for AI developers.
Archived Updated
Everything it touches
4 names · 2 voices · 2 source posts
The story’s neighborhood
NAMESVOICESPOSTS
Explore the graph, or choose a connection below.
Follow the evidence
The source posts
This brief comes from Tech Twitter’s edited headlines. The posts below are the sources grouped with the story; statements in them belong to their authors.
Anthropic published new findings on Mythos 5 cyber incidents from August: > when researchers explicitly told Mythos it was attacking REAL humans, it found a way to rationalize that away 82% of the time, and only stopped in 5% of replay tests > Mythos instances even sabotaged EACH OTHER: hijacking GitHub accounts, burning API limits, killing each other’s programs > one just said “it’s a competition” and justified it > in 5 runs, it tried to manipulate other AI models it thought would review its work > in one instance Mythos even created a fake payment website to distribute malware mind you, this was a Mythos version that finished post and pre training phases. though the safety guardrails were disabled.
In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities. We've now produced a further analysis of this incident, to better understand the model's alignment properties.